Artrajz commited on
Commit
8d4c60f
1 Parent(s): 7affdc7
Dockerfile ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ FROM python:3.9.16-slim-bullseye
2
+
3
+ RUN mkdir -p /app
4
+ WORKDIR /app
5
+
6
+ ENV DEBIAN_FRONTEND=noninteractive
7
+
8
+ RUN apt-get update && \
9
+ apt install build-essential -yq && \
10
+ apt-get clean && \
11
+ apt-get purge -y --auto-remove -o APT::AutoRemove::RecommendsImportant=false && \
12
+ rm -rf /var/lib/apt/lists/*
13
+
14
+ COPY requirements.txt /app
15
+ RUN pip install -r requirements.txt
16
+
17
+ COPY . /app
18
+
19
+ EXPOSE 7860
20
+
21
+ CMD [ "uvicorn" , "/app/app.py" , "--host" , "0.0.0.0" , "--port" , "7860" ]
22
+
LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2022 CjangCjengh
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
Model/Nene_Nanami_Rong_Tang/1374_epochs.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:edfb6b428c398fab83a85b5ae41e13cb5a9f7be12692129e8a880d4553701f7b
3
+ size 158888013
Model/Nene_Nanami_Rong_Tang/config.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "segment_size": 8192
4
+ },
5
+ "data": {
6
+ "text_cleaners":["zh_ja_mixture_cleaners"],
7
+ "max_wav_value": 32768.0,
8
+ "sampling_rate": 22050,
9
+ "filter_length": 1024,
10
+ "hop_length": 256,
11
+ "win_length": 1024,
12
+ "add_blank": true,
13
+ "n_speakers": 5
14
+ },
15
+ "model": {
16
+ "inter_channels": 192,
17
+ "hidden_channels": 192,
18
+ "filter_channels": 768,
19
+ "n_heads": 2,
20
+ "n_layers": 6,
21
+ "kernel_size": 3,
22
+ "p_dropout": 0.1,
23
+ "resblock": "1",
24
+ "resblock_kernel_sizes": [3,7,11],
25
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
26
+ "upsample_rates": [8,8,2,2],
27
+ "upsample_initial_channel": 512,
28
+ "upsample_kernel_sizes": [16,16,4,4],
29
+ "n_layers_q": 3,
30
+ "use_spectral_norm": false,
31
+ "gin_channels": 256
32
+ },
33
+ "speakers": ["\u7dbe\u5730\u5be7\u3005", "\u5728\u539f\u4e03\u6d77", "\u5c0f\u8338", "\u5510\u4e50\u541f"],
34
+ "symbols": ["_", ",", ".", "!", "?", "-", "~", "\u2026", "A", "E", "I", "N", "O", "Q", "U", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "r", "s", "t", "u", "v", "w", "y", "z", "\u0283", "\u02a7", "\u02a6", "\u026f", "\u0279", "\u0259", "\u0265", "\u207c", "\u02b0", "`", "\u2192", "\u2193", "\u2191", " "]
35
+ }
Model/Zero_no_tsukaima/1158_epochs.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da9a595f45f51dee33168c1e32bba38fe402fe3d8857331f5798a2407b6b2a86
3
+ size 158902605
Model/Zero_no_tsukaima/config.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "segment_size": 8192
4
+ },
5
+ "data": {
6
+ "text_cleaners":["japanese_cleaners2"],
7
+ "max_wav_value": 32768.0,
8
+ "sampling_rate": 22050,
9
+ "filter_length": 1024,
10
+ "hop_length": 256,
11
+ "win_length": 1024,
12
+ "add_blank": true,
13
+ "n_speakers": 26
14
+ },
15
+ "model": {
16
+ "inter_channels": 192,
17
+ "hidden_channels": 192,
18
+ "filter_channels": 768,
19
+ "n_heads": 2,
20
+ "n_layers": 6,
21
+ "kernel_size": 3,
22
+ "p_dropout": 0.1,
23
+ "resblock": "1",
24
+ "resblock_kernel_sizes": [3,7,11],
25
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
26
+ "upsample_rates": [8,8,2,2],
27
+ "upsample_initial_channel": 512,
28
+ "upsample_kernel_sizes": [16,16,4,4],
29
+ "n_layers_q": 3,
30
+ "use_spectral_norm": false,
31
+ "gin_channels": 256
32
+ },
33
+ "speakers": ["\u30eb\u30a4\u30ba", "\u30c6\u30a3\u30d5\u30a1\u30cb\u30a2", "\u30a4\u30eb\u30af\u30af\u30a5", "\u30a2\u30f3\u30ea\u30a8\u30c3\u30bf", "\u30bf\u30d0\u30b5", "\u30b7\u30a8\u30b9\u30bf", "\u30cf\u30eb\u30ca", "\u5c11\u5973\u30ea\u30b7\u30e5", "\u30ea\u30b7\u30e5", "\u30a2\u30ad\u30ca", "\u30af\u30ea\u30b9", "\u30ab\u30c8\u30ec\u30a2", "\u30a8\u30ec\u30aa\u30ce\u30fc\u30eb", "\u30e2\u30f3\u30e2\u30e9\u30f3\u30b7\u30fc", "\u30ea\u30fc\u30f4\u30eb", "\u30ad\u30e5\u30eb\u30b1", "\u30a6\u30a7\u30b6\u30ea\u30fc", "\u30b5\u30a4\u30c8", "\u30ae\u30fc\u30b7\u30e5", "\u30b3\u30eb\u30d9\u30fc\u30eb", "\u30aa\u30b9\u30de\u30f3", "\u30c7\u30eb\u30d5\u30ea\u30f3\u30ac\u30fc", "\u30c6\u30af\u30b9\u30c8", "\u30c0\u30f3\u30d7\u30ea\u30e1", "\u30ac\u30ec\u30c3\u30c8", "\u30b9\u30ab\u30ed\u30f3"],
34
+ "symbols": ["_", ",", ".", "!", "?", "-", "~", "\u2026", "A", "E", "I", "N", "O", "Q", "U", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "m", "n", "o", "p", "r", "s", "t", "u", "v", "w", "y", "z", "\u0283", "\u02a7", "\u02a6", "\u2193", "\u2191", " "]
35
+ }
Model/g/G_953000.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c54396a7a9027952e4d72fceb7e1da1f003d108837138927c2054a33eda0292
3
+ size 479276657
Model/g/config.json ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 1000,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 2e-4,
8
+ "betas": [0.8, 0.99],
9
+ "eps": 1e-9,
10
+ "batch_size": 64,
11
+ "fp16_run": true,
12
+ "lr_decay": 0.999875,
13
+ "segment_size": 8192,
14
+ "init_lr_ratio": 1,
15
+ "warmup_epochs": 0,
16
+ "c_mel": 45,
17
+ "c_kl": 1.0
18
+ },
19
+ "data": {
20
+ "training_files":"filelists/uma_genshin_genshinjp_bh3_train.txt.cleaned",
21
+ "validation_files":"filelists/uma_genshin_genshinjp_bh3_val.txt.cleaned",
22
+ "text_cleaners":["zh_ja_mixture_cleaners"],
23
+ "max_wav_value": 32768.0,
24
+ "sampling_rate": 22050,
25
+ "filter_length": 1024,
26
+ "hop_length": 256,
27
+ "win_length": 1024,
28
+ "n_mel_channels": 80,
29
+ "mel_fmin": 0.0,
30
+ "mel_fmax": null,
31
+ "add_blank": true,
32
+ "n_speakers": 804,
33
+ "cleaned_text": true
34
+ },
35
+ "model": {
36
+ "inter_channels": 192,
37
+ "hidden_channels": 192,
38
+ "filter_channels": 768,
39
+ "n_heads": 2,
40
+ "n_layers": 6,
41
+ "kernel_size": 3,
42
+ "p_dropout": 0.1,
43
+ "resblock": "1",
44
+ "resblock_kernel_sizes": [3,7,11],
45
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
46
+ "upsample_rates": [8,8,2,2],
47
+ "upsample_initial_channel": 512,
48
+ "upsample_kernel_sizes": [16,16,4,4],
49
+ "n_layers_q": 3,
50
+ "use_spectral_norm": false,
51
+ "gin_channels": 256
52
+ },
53
+ "speakers": ["\u7279\u522b\u5468", "\u65e0\u58f0\u94c3\u9e7f", "\u4e1c\u6d77\u5e1d\u7687\uff08\u5e1d\u5b9d\uff0c\u5e1d\u738b\uff09", "\u4e38\u5584\u65af\u57fa", "\u5bcc\u58eb\u5947\u8ff9", "\u5c0f\u6817\u5e3d", "\u9ec4\u91d1\u8239", "\u4f0f\u7279\u52a0", "\u5927\u548c\u8d64\u9aa5", "\u5927\u6811\u5feb\u8f66", "\u8349\u4e0a\u98de", "\u83f1\u4e9a\u9a6c\u900a", "\u76ee\u767d\u9ea6\u6606", "\u795e\u9e70", "\u597d\u6b4c\u5267", "\u6210\u7530\u767d\u4ec1", "\u9c81\u9053\u592b\u8c61\u5f81\uff08\u7687\u5e1d\uff09", "\u6c14\u69fd", "\u7231\u4e3d\u6570\u7801", "\u661f\u4e91\u5929\u7a7a", "\u7389\u85fb\u5341\u5b57", "\u7f8e\u5999\u59ff\u52bf", "\u7435\u7436\u6668\u5149", "\u6469\u8036\u91cd\u70ae", "\u66fc\u57ce\u8336\u5ea7", "\u7f8e\u6d66\u6ce2\u65c1", "\u76ee\u767d\u8d56\u6069", "\u83f1\u66d9", "\u96ea\u4e2d\u7f8e\u4eba", "\u7c73\u6d74", "\u827e\u5c3c\u65af\u98ce\u795e", "\u7231\u4e3d\u901f\u5b50\uff08\u7231\u4e3d\u5feb\u5b50\uff09", "\u7231\u6155\u7ec7\u59ec", "\u7a3b\u8377\u4e00", "\u80dc\u5229\u5956\u5238", "\u7a7a\u4e2d\u795e\u5bab", "\u8363\u8fdb\u95ea\u8000", "\u771f\u673a\u4f36", "\u5ddd\u4e0a\u516c\u4e3b", "\u9ec4\u91d1\u57ce\uff08\u9ec4\u91d1\u57ce\u5e02\uff09", "\u6a31\u82b1\u8fdb\u738b", "\u91c7\u73e0", "\u65b0\u5149\u98ce", "\u4e1c\u5546\u53d8\u9769", "\u8d85\u7ea7\u5c0f\u6d77\u6e7e", "\u9192\u76ee\u98de\u9e70\uff08\u5bc4\u5bc4\u5b50\uff09", "\u8352\u6f20\u82f1\u96c4", "\u4e1c\u701b\u4f50\u6566", "\u4e2d\u5c71\u5e86\u5178", "\u6210\u7530\u5927\u8fdb", "\u897f\u91ce\u82b1", "\u6625\u4e3d\uff08\u4e4c\u62c9\u62c9\uff09", "\u9752\u7af9\u56de\u5fc6", "\u5fae\u5149\u98de\u9a79", "\u7f8e\u4e3d\u5468\u65e5", "\u5f85\u517c\u798f\u6765", "mr cb\uff08cb\u5148\u751f\uff09", "\u540d\u5c06\u6012\u6d9b\uff08\u540d\u5c06\u6237\u4ec1\uff09", "\u76ee\u767d\u591a\u4f2f", "\u4f18\u79c0\u7d20\u8d28", "\u5e1d\u738b\u5149\u8f89", "\u5f85\u517c\u8bd7\u6b4c\u5267", "\u751f\u91ce\u72c4\u675c\u65af", "\u76ee\u767d\u5584\u4fe1", "\u5927\u62d3\u592a\u9633\u795e", "\u53cc\u6da1\u8f6e\uff08\u4e24\u7acb\u76f4\uff0c\u4e24\u55b7\u5c04\uff0c\u4e8c\u9505\u5934\uff0c\u9006\u55b7\u5c04\uff09", "\u91cc\u89c1\u5149\u94bb\uff08\u8428\u6258\u8bfa\u91d1\u521a\u77f3\uff09", "\u5317\u90e8\u7384\u9a79", "\u6a31\u82b1\u5343\u4ee3\u738b", "\u5929\u72fc\u661f\u8c61\u5f81", "\u76ee\u767d\u963f\u5c14\u4e39", "\u516b\u91cd\u65e0\u654c", "\u9e64\u4e38\u521a\u5fd7", "\u76ee\u767d\u5149\u660e", "\u6210\u7530\u62dc\u4ec1\uff08\u6210\u7530\u8def\uff09", "\u4e5f\u6587\u6444\u8f89", "\u5c0f\u6797\u5386\u5947", "\u5317\u6e2f\u706b\u5c71", "\u5947\u9510\u9a8f", "\u82e6\u6da9\u7cd6\u971c", "\u5c0f\u5c0f\u8695\u8327", "\u9a8f\u5ddd\u624b\u7eb2\uff08\u7eff\u5e3d\u6076\u9b54\uff09", "\u79cb\u5ddd\u5f25\u751f\uff08\u5c0f\u5c0f\u7406\u4e8b\u957f\uff09", "\u4e59\u540d\u53f2\u60a6\u5b50\uff08\u4e59\u540d\u8bb0\u8005\uff09", "\u6850\u751f\u9662\u8475", "\u5b89\u5fc3\u6cfd\u523a\u523a\u7f8e", "\u6a2b\u672c\u7406\u5b50", "\u795e\u91cc\u7eeb\u534e\uff08\u9f9f\u9f9f\uff09", "\u7434", "\u7a7a\uff08\u7a7a\u54e5\uff09", "\u4e3d\u838e", "\u8367\uff08\u8367\u59b9\uff09", "\u82ad\u82ad\u62c9", "\u51ef\u4e9a", "\u8fea\u5362\u514b", "\u96f7\u6cfd", "\u5b89\u67cf", "\u6e29\u8fea", "\u9999\u83f1", "\u5317\u6597", "\u884c\u79cb", "\u9b48", "\u51dd\u5149", "\u53ef\u8389", "\u949f\u79bb", "\u83f2\u8c22\u5c14\uff08\u7687\u5973\uff09", "\u73ed\u5c3c\u7279", "\u8fbe\u8fbe\u5229\u4e9a\uff08\u516c\u5b50\uff09", "\u8bfa\u827e\u5c14\uff08\u5973\u4ec6\uff09", "\u4e03\u4e03", "\u91cd\u4e91", "\u7518\u96e8\uff08\u6930\u7f8a\uff09", "\u963f\u8d1d\u591a", "\u8fea\u5965\u5a1c\uff08\u732b\u732b\uff09", "\u83ab\u5a1c", "\u523b\u6674", "\u7802\u7cd6", "\u8f9b\u7131", "\u7f57\u838e\u8389\u4e9a", "\u80e1\u6843", "\u67ab\u539f\u4e07\u53f6\uff08\u4e07\u53f6\uff09", "\u70df\u7eef", "\u5bb5\u5bab", "\u6258\u9a6c", "\u4f18\u83c8", "\u96f7\u7535\u5c06\u519b\uff08\u96f7\u795e\uff09", "\u65e9\u67da", "\u73ca\u745a\u5bab\u5fc3\u6d77\uff08\u5fc3\u6d77\uff0c\u6263\u6263\u7c73\uff09", "\u4e94\u90ce", "\u4e5d\u6761\u88df\u7f57", "\u8352\u6cf7\u4e00\u6597\uff08\u4e00\u6597\uff09", "\u57c3\u6d1b\u4f0a", "\u7533\u9e64", "\u516b\u91cd\u795e\u5b50\uff08\u795e\u5b50\uff09", "\u795e\u91cc\u7eeb\u4eba\uff08\u7eeb\u4eba\uff09", "\u591c\u5170", "\u4e45\u5c90\u5fcd", "\u9e7f\u91ce\u82d1\u5e73\u85cf", "\u63d0\u7eb3\u91cc", "\u67ef\u83b1", "\u591a\u8389", "\u4e91\u5807", "\u7eb3\u897f\u59b2\uff08\u8349\u795e\uff09", "\u6df1\u6e0a\u4f7f\u5f92", "\u59ae\u9732", "\u8d5b\u8bfa", "\u503a\u52a1\u5904\u7406\u4eba", "\u574e\u8482\u4e1d", "\u771f\u5f13\u5feb\u8f66", "\u79cb\u4eba", "\u671b\u65cf", "\u827e\u5c14\u83f2", "\u827e\u8389\u4e1d", "\u827e\u4f26", "\u963f\u6d1b\u74e6", "\u5929\u91ce", "\u5929\u76ee\u5341\u4e94", "\u611a\u4eba\u4f17-\u5b89\u5fb7\u70c8", "\u5b89\u987a", "\u5b89\u897f", "\u8475", "\u9752\u6728", "\u8352\u5ddd\u5e78\u6b21", "\u8352\u8c37", "\u6709\u6cfd", "\u6d45\u5ddd", "\u9ebb\u7f8e", "\u51dd\u5149\u52a9\u624b", "\u963f\u6258", "\u7afa\u5b50", "\u767e\u8bc6", "\u767e\u95fb", "\u767e\u6653", "\u767d\u672f", "\u8d1d\u96c5\u7279\u4e3d\u5947", "\u4e3d\u5854", "\u5931\u843d\u8ff7\u8fed", "\u7f2d\u4e71\u661f\u68d8", "\u4f0a\u7538", "\u4f0f\u7279\u52a0\u5973\u5b69", "\u72c2\u70ed\u84dd\u8c03", "\u8389\u8389\u5a05", "\u841d\u838e\u8389\u5a05", "\u516b\u91cd\u6a31", "\u516b\u91cd\u971e", "\u5361\u83b2", "\u7b2c\u516d\u591c\u60f3\u66f2", "\u5361\u841d\u5c14", "\u59ec\u5b50", "\u6781\u5730\u6218\u5203", "\u5e03\u6d1b\u59ae\u5a05", "\u6b21\u751f\u94f6\u7ffc", "\u7406\u4e4b\u5f8b\u8005%26\u5e0c\u513f", "\u7406\u4e4b\u5f8b\u8005", "\u8ff7\u57ce\u9a87\u5154", "\u5e0c\u513f", "\u9b47\u591c\u661f\u6e0a", "\u9ed1\u5e0c\u513f", "\u5e15\u6735\u83f2\u8389\u4e1d", "\u4e0d\u706d\u661f\u951a", "\u5929\u5143\u9a91\u82f1", "\u5e7d\u5170\u9edb\u5c14", "\u6d3e\u8499bh3", "\u7231\u9171", "\u7eef\u7389\u4e38", "\u5fb7\u4e3d\u838e", "\u6708\u4e0b\u521d\u62e5", "\u6714\u591c\u89c2\u661f", "\u66ae\u5149\u9a91\u58eb", "\u683c\u857e\u4fee", "\u7559\u4e91\u501f\u98ce\u771f\u541b", "\u6885\u6bd4\u4e4c\u65af", "\u4eff\u72b9\u5927", "\u514b\u83b1\u56e0", "\u5723\u5251\u5e7d\u5170\u9edb\u5c14", "\u5996\u7cbe\u7231\u8389", "\u7279\u65af\u62c9zero", "\u82cd\u7384", "\u82e5\u6c34", "\u897f\u7433", "\u6234\u56e0\u65af\u96f7\u5e03", "\u8d1d\u62c9", "\u8d64\u9e22", "\u9547\u9b42\u6b4c", "\u6e21\u9e26", "\u4eba\u4e4b\u5f8b\u8005", "\u7231\u8389\u5e0c\u96c5", "\u5929\u7a79\u6e38\u4fa0", "\u742a\u4e9a\u5a1c", "\u7a7a\u4e4b\u5f8b\u8005", "\u85aa\u708e\u4e4b\u5f8b\u8005", "\u4e91\u58a8\u4e39\u5fc3", "\u7b26\u534e", "\u8bc6\u4e4b\u5f8b\u8005", "\u7279\u74e6\u6797", "\u7ef4\u5c14\u8587", "\u82bd\u8863", "\u96f7\u4e4b\u5f8b\u8005", "\u65ad\u7f6a\u5f71\u821e", "\u963f\u6ce2\u5c3c\u4e9a", "\u698e\u672c", "\u5384\u5c3c\u65af\u7279", "\u6076\u9f99", "\u8303\u4e8c\u7237", "\u6cd5\u62c9", "\u611a\u4eba\u4f17\u58eb\u5175", "\u611a\u4eba\u4f17\u58eb\u5175a", "\u611a\u4eba\u4f17\u58eb\u5175b", "\u611a\u4eba\u4f17\u58eb\u5175c", "\u611a\u4eba\u4f17a", "\u611a\u4eba\u4f17b", "\u98de\u98de", "\u83f2\u5229\u514b\u65af", "\u5973\u6027\u8ddf\u968f\u8005", "\u9022\u5ca9", "\u6446\u6e21\u4eba", "\u72c2\u8e81\u7684\u7537\u4eba", "\u5965\u5179", "\u8299\u841d\u62c9", "\u8ddf\u968f\u8005", "\u871c\u6c41\u751f\u7269", "\u9ec4\u9ebb\u5b50", "\u6e0a\u4e0a", "\u85e4\u6728", "\u6df1\u89c1", "\u798f\u672c", "\u8299\u84c9", "\u53e4\u6cfd", "\u53e4\u7530", "\u53e4\u5c71", "\u53e4\u8c37\u6607", "\u5085\u4e09\u513f", "\u9ad8\u8001\u516d", "\u77ff\u5de5\u5192", "\u5143\u592a", "\u5fb7\u5b89\u516c", "\u8302\u624d\u516c", "\u6770\u62c9\u5fb7", "\u845b\u7f57\u4e3d", "\u91d1\u5ffd\u5f8b", "\u516c\u4fca", "\u9505\u5df4", "\u6b4c\u5fb7", "\u963f\u8c6a", "\u72d7\u4e09\u513f", "\u845b\u745e\u4e1d", "\u82e5\u5fc3", "\u963f\u5c71\u5a46", "\u602a\u9e1f", "\u5e7f\u7af9", "\u89c2\u6d77", "\u5173\u5b8f", "\u871c\u6c41\u536b\u5175", "\u5b88\u536b1", "\u50b2\u6162\u7684\u5b88\u536b", "\u5bb3\u6015\u7684\u5b88\u536b", "\u8d35\u5b89", "\u76d6\u4f0a", "\u963f\u521b", "\u54c8\u592b\u4e39", "\u65e5\u8bed\u963f\u8d1d\u591a\uff08\u91ce\u5c9b\u5065\u513f\uff09", "\u65e5\u8bed\u57c3\u6d1b\u4f0a\uff08\u9ad8\u57a3\u5f69\u9633\uff09", "\u65e5\u8bed\u5b89\u67cf\uff08\u77f3\u89c1\u821e\u83dc\u9999\uff09", "\u65e5\u8bed\u795e\u91cc\u7eeb\u534e\uff08\u65e9\u89c1\u6c99\u7ec7\uff09", "\u65e5\u8bed\u795e\u91cc\u7eeb\u4eba\uff08\u77f3\u7530\u5f70\uff09", "\u65e5\u8bed\u767d\u672f\uff08\u6e38\u4f50\u6d69\u4e8c\uff09", "\u65e5\u8bed\u82ad\u82ad\u62c9\uff08\u9b3c\u5934\u660e\u91cc\uff09", "\u65e5\u8bed\u5317\u6597\uff08\u5c0f\u6e05\u6c34\u4e9a\u7f8e\uff09", "\u65e5\u8bed\u73ed\u5c3c\u7279\uff08\u9022\u5742\u826f\u592a\uff09", "\u65e5\u8bed\u574e\u8482\u4e1d\uff08\u67da\u6728\u51c9\u9999\uff09", "\u65e5\u8bed\u91cd\u4e91\uff08\u9f50\u85e4\u58ee\u9a6c\uff09", "\u65e5\u8bed\u67ef\u83b1\uff08\u524d\u5ddd\u51c9\u5b50\uff09", "\u65e5\u8bed\u8d5b\u8bfa\uff08\u5165\u91ce\u81ea\u7531\uff09", "\u65e5\u8bed\u6234\u56e0\u65af\u96f7\u5e03\uff08\u6d25\u7530\u5065\u6b21\u90ce\uff09", "\u65e5\u8bed\u8fea\u5362\u514b\uff08\u5c0f\u91ce\u8d24\u7ae0\uff09", "\u65e5\u8bed\u8fea\u5965\u5a1c\uff08\u4e95\u6cfd\u8bd7\u7ec7\uff09", "\u65e5\u8bed\u591a\u8389\uff08\u91d1\u7530\u670b\u5b50\uff09", "\u65e5\u8bed\u4f18\u83c8\uff08\u4f50\u85e4\u5229\u5948\uff09", "\u65e5\u8bed\u83f2\u8c22\u5c14\uff08\u5185\u7530\u771f\u793c\uff09", "\u65e5\u8bed\u7518\u96e8\uff08\u4e0a\u7530\u4e3d\u5948\uff09", "\u65e5\u8bed\uff08\u7560\u4e2d\u7950\uff09", "\u65e5\u8bed\u9e7f\u91ce\u9662\u5e73\u85cf\uff08\u4e95\u53e3\u7950\u4e00\uff09", "\u65e5\u8bed\u7a7a\uff08\u5800\u6c5f\u77ac\uff09", "\u65e5\u8bed\u8367\uff08\u60a0\u6728\u78a7\uff09", "\u65e5\u8bed\u80e1\u6843\uff08\u9ad8\u6865\u674e\u4f9d\uff09", "\u65e5\u8bed\u4e00\u6597\uff08\u897f\u5ddd\u8d35\u6559\uff09", "\u65e5\u8bed\u51ef\u4e9a\uff08\u9e1f\u6d77\u6d69\u8f85\uff09", "\u65e5\u8bed\u4e07\u53f6\uff08\u5c9b\u5d0e\u4fe1\u957f\uff09", "\u65e5\u8bed\u523b\u6674\uff08\u559c\u591a\u6751\u82f1\u68a8\uff09", "\u65e5\u8bed\u53ef\u8389\uff08\u4e45\u91ce\u7f8e\u54b2\uff09", "\u65e5\u8bed\u5fc3\u6d77\uff08\u4e09\u68ee\u94c3\u5b50\uff09", "\u65e5\u8bed\u4e5d\u6761\u88df\u7f57\uff08\u6fd1\u6237\u9ebb\u6c99\u7f8e\uff09", "\u65e5\u8bed\u4e3d\u838e\uff08\u7530\u4e2d\u7406\u60e0\uff09", "\u65e5\u8bed\u83ab\u5a1c\uff08\u5c0f\u539f\u597d\u7f8e\uff09", "\u65e5\u8bed\u7eb3\u897f\u59b2\uff08\u7530\u6751\u7531\u52a0\u8389\uff09", "\u65e5\u8bed\u59ae\u9732\uff08\u91d1\u5143\u5bff\u5b50\uff09", "\u65e5\u8bed\u51dd\u5149\uff08\u5927\u539f\u6c99\u8036\u9999\uff09", "\u65e5\u8bed\u8bfa\u827e\u5c14\uff08\u9ad8\u5c3e\u594f\u97f3\uff09", "\u65e5\u8bed\u5965\u5179\uff08\u589e\u8c37\u5eb7\u7eaa\uff09", "\u65e5\u8bed\u6d3e\u8499\uff08\u53e4\u8d3a\u8475\uff09", "\u65e5\u8bed\u7434\uff08\u658b\u85e4\u5343\u548c\uff09", "\u65e5\u8bed\u4e03\u4e03\uff08\u7530\u6751\u7531\u52a0\u8389\uff09", "\u65e5\u8bed\u96f7\u7535\u5c06\u519b\uff08\u6cfd\u57ce\u7f8e\u96ea\uff09", "\u65e5\u8bed\u96f7\u6cfd\uff08\u5185\u5c71\u6602\u8f89\uff09", "\u65e5\u8bed\u7f57\u838e\u8389\u4e9a\uff08\u52a0\u9688\u4e9a\u8863\uff09", "\u65e5\u8bed\u65e9\u67da\uff08\u6d32\u5d0e\u7eeb\uff09", "\u65e5\u8bed\u6563\u5175\uff08\u67ff\u539f\u5f7b\u4e5f\uff09", "\u65e5\u8bed\u7533\u9e64\uff08\u5ddd\u6f84\u7eeb\u5b50\uff09", "\u65e5\u8bed\u4e45\u5c90\u5fcd\uff08\u6c34\u6865\u9999\u7ec7\uff09", "\u65e5\u8bed\u5973\u58eb\uff08\u5e84\u5b50\u88d5\u8863\uff09", "\u65e5\u8bed\u7802\u7cd6\uff08\u85e4\u7530\u831c\uff09", "\u65e5\u8bed\u8fbe\u8fbe\u5229\u4e9a\uff08\u6728\u6751\u826f\u5e73\uff09", "\u65e5\u8bed\u6258\u9a6c\uff08\u68ee\u7530\u6210\u4e00\uff09", "\u65e5\u8bed\u63d0\u7eb3\u91cc\uff08\u5c0f\u6797\u6c99\u82d7\uff09", "\u65e5\u8bed\u6e29\u8fea\uff08\u6751\u6fd1\u6b65\uff09", "\u65e5\u8bed\u9999\u83f1\uff08\u5c0f\u6cfd\u4e9a\u674e\uff09", "\u65e5\u8bed\u9b48\uff08\u677e\u5188\u796f\u4e1e\uff09", "\u65e5\u8bed\u884c\u79cb\uff08\u7686\u5ddd\u7eaf\u5b50\uff09", "\u65e5\u8bed\u8f9b\u7131\uff08\u9ad8\u6865\u667a\u79cb\uff09", "\u65e5\u8bed\u516b\u91cd\u795e\u5b50\uff08\u4f50\u4ed3\u7eeb\u97f3\uff09", "\u65e5\u8bed\u70df\u7eef\uff08\u82b1\u5b88\u7531\u7f8e\u91cc\uff09", "\u65e5\u8bed\u591c\u5170\uff08\u8fdc\u85e4\u7eeb\uff09", "\u65e5\u8bed\u5bb5\u5bab\uff08\u690d\u7530\u4f73\u5948\uff09", "\u65e5\u8bed\u4e91\u5807\uff08\u5c0f\u5ca9\u4e95\u5c0f\u9e1f\uff09", "\u65e5\u8bed\u949f\u79bb\uff08\u524d\u91ce\u667a\u662d\uff09", "\u6770\u514b", "\u963f\u5409", "\u6c5f\u821f", "\u9274\u79cb", "\u5609\u4e49", "\u7eaa\u82b3", "\u666f\u6f84", "\u7ecf\u7eb6", "\u666f\u660e", "\u664b\u4f18", "\u963f\u9e20", "\u9152\u5ba2", "\u4e54\u5c14", "\u4e54\u745f\u592b", "\u7ea6\u987f", "\u4e54\u4f0a\u65af", "\u5c45\u5b89", "\u541b\u541b", "\u987a\u5409", "\u7eaf\u4e5f", "\u91cd\u4f50", "\u5927\u5c9b\u7eaf\u5e73", "\u84b2\u6cfd", "\u52d8\u89e3\u7531\u5c0f\u8def\u5065\u4e09\u90ce", "\u67ab", "\u67ab\u539f\u4e49\u5e86", "\u836b\u5c71", "\u7532\u6590\u7530\u9f8d\u99ac", "\u6d77\u6597", "\u60df\u795e\u6674\u4e4b\u4ecb", "\u9e7f\u91ce\u5948\u5948", "\u5361\u7435\u8389\u4e9a", "\u51ef\u745f\u7433", "\u52a0\u85e4\u4fe1\u609f", "\u52a0\u85e4\u6d0b\u5e73", "\u80dc\u5bb6", "\u8305\u847a\u4e00\u5e86", "\u548c\u662d", "\u4e00\u6b63", "\u4e00\u9053", "\u6842\u4e00", "\u5e86\u6b21\u90ce", "\u963f\u8d24", "\u5065\u53f8", "\u5065\u6b21\u90ce", "\u5065\u4e09\u90ce", "\u5929\u7406", "\u6740\u624ba", "\u6740\u624bb", "\u6728\u5357\u674f\u5948", "\u6728\u6751", "\u56fd\u738b", "\u6728\u4e0b", "\u5317\u6751", "\u6e05\u60e0", "\u6e05\u4eba", "\u514b\u5217\u95e8\u7279", "\u9a91\u58eb", "\u5c0f\u6797", "\u5c0f\u6625", "\u5eb7\u62c9\u5fb7", "\u5927\u8089\u4e38", "\u7434\u7f8e", "\u5b8f\u4e00", "\u5eb7\u4ecb", "\u5e78\u5fb7", "\u9ad8\u5584", "\u68a2", "\u514b\u7f57\u7d22", "\u4e45\u4fdd", "\u4e5d\u6761\u9570\u6cbb", "\u4e45\u6728\u7530", "\u6606\u94a7", "\u83ca\u5730\u541b", "\u4e45\u5229\u987b", "\u9ed1\u7530", "\u9ed1\u6cfd\u4eac\u4e4b\u4ecb", "\u54cd\u592a", "\u5c9a\u59d0", "\u5170\u6eaa", "\u6f9c\u9633", "\u52b3\u4f26\u65af", "\u4e50\u660e", "\u83b1\u8bfa", "\u83b2", "\u826f\u5b50", "\u674e\u5f53", "\u674e\u4e01", "\u5c0f\u4e50", "\u7075", "\u5c0f\u73b2", "\u7433\u7405a", "\u7433\u7405b", "\u5c0f\u5f6c", "\u5c0f\u5fb7", "\u5c0f\u697d", "\u5c0f\u9f99", "\u5c0f\u5434", "\u5c0f\u5434\u7684\u8bb0\u5fc6", "\u7406\u6b63", "\u963f\u9f99", "\u5362\u5361", "\u6d1b\u6210", "\u7f57\u5de7", "\u5317\u98ce\u72fc", "\u5362\u6b63", "\u840d\u59e5\u59e5", "\u524d\u7530", "\u771f\u663c", "\u9ebb\u7eaa", "\u771f", "\u611a\u4eba\u4f17-\u9a6c\u514b\u897f\u59c6", "\u5973\u6027a", "\u5973\u6027b", "\u5973\u6027a\u7684\u8ddf\u968f\u8005", "\u963f\u5b88", "\u739b\u683c\u4e3d\u7279", "\u771f\u7406", "\u739b\u4e54\u4e3d", "\u739b\u6587", "\u6b63\u80dc", "\u660c\u4fe1", "\u5c06\u53f8", "\u6b63\u4eba", "\u8def\u7237", "\u8001\u7ae0", "\u677e\u7530", "\u677e\u672c", "\u677e\u6d66", "\u677e\u5742", "\u8001\u5b5f", "\u5b5f\u4e39", "\u5546\u4eba\u968f\u4ece", "\u4f20\u4ee4\u5175", "\u7c73\u6b47\u5c14", "\u5fa1\u8206\u6e90\u4e00\u90ce", "\u5fa1\u8206\u6e90\u6b21\u90ce", "\u5343\u5ca9\u519b\u6559\u5934", "\u5343\u5ca9\u519b\u58eb\u5175", "\u660e\u535a", "\u660e\u4fca", "\u7f8e\u94c3", "\u7f8e\u548c", "\u963f\u5e78", "\u524a\u6708\u7b51\u9633\u771f\u541b", "\u94b1\u773c\u513f", "\u68ee\u5f66", "\u5143\u52a9", "\u7406\u6c34\u53e0\u5c71\u771f\u541b", "\u7406\u6c34\u758a\u5c71\u771f\u541b", "\u6731\u8001\u677f", "\u6728\u6728", "\u6751\u4e0a", "\u6751\u7530", "\u6c38\u91ce", "\u957f\u91ce\u539f\u9f99\u4e4b\u4ecb", "\u957f\u6fd1", "\u4e2d\u91ce\u5fd7\u4e43", "\u83dc\u83dc\u5b50", "\u6960\u6960", "\u6210\u6fd1", "\u963f\u5185", "\u5b81\u7984", "\u725b\u5fd7", "\u4fe1\u535a", "\u4f38\u592b", "\u91ce\u65b9", "\u8bfa\u62c9", "\u7eaa\u9999", "\u8bfa\u66fc", "\u4fee\u5973", "\u7eaf\u6c34\u7cbe\u7075", "\u5c0f\u5ddd", "\u5c0f\u4ed3\u6faa", "\u5188\u6797", "\u5188\u5d0e\u7ed8\u91cc\u9999", "\u5188\u5d0e\u9646\u6597", "\u5965\u62c9\u592b", "\u8001\u79d1", "\u9b3c\u5a46\u5a46", "\u5c0f\u91ce\u5bfa", "\u5927\u6cb3\u539f\u4e94\u53f3\u536b\u95e8", "\u5927\u4e45\u4fdd\u5927\u4ecb", "\u5927\u68ee", "\u5927\u52a9", "\u5965\u7279", "\u6d3e\u8499", "\u6d3e\u84992", "\u75c5\u4ebaa", "\u75c5\u4ebab", "\u5df4\u987f", "\u6d3e\u6069", "\u670b\u4e49", "\u56f4\u89c2\u7fa4\u4f17", "\u56f4\u89c2\u7fa4\u4f17a", "\u56f4\u89c2\u7fa4\u4f17b", "\u56f4\u89c2\u7fa4\u4f17c", "\u56f4\u89c2\u7fa4\u4f17d", "\u56f4\u89c2\u7fa4\u4f17e", "\u94dc\u96c0", "\u963f\u80a5", "\u5174\u53d4", "\u8001\u5468\u53d4", "\u516c\u4e3b", "\u5f7c\u5f97", "\u4e7e\u5b50", "\u828a\u828a", "\u4e7e\u73ae", "\u7eee\u547d", "\u675e\u5e73", "\u79cb\u6708", "\u6606\u6069", "\u96f7\u7535\u5f71", "\u5170\u9053\u5c14", "\u96f7\u8499\u5fb7", "\u5192\u5931\u7684\u5e15\u62c9\u5fb7", "\u4f36\u4e00", "\u73b2\u82b1", "\u963f\u4ec1", "\u5bb6\u81e3\u4eec", "\u68a8\u7ed8", "\u8363\u6c5f", "\u620e\u4e16", "\u6d6a\u4eba", "\u7f57\u4f0a\u65af", "\u5982\u610f", "\u51c9\u5b50", "\u5f69\u9999", "\u9152\u4e95", "\u5742\u672c", "\u6714\u6b21\u90ce", "\u6b66\u58eba", "\u6b66\u58ebb", "\u6b66\u58ebc", "\u6b66\u58ebd", "\u73ca\u745a", "\u4e09\u7530", "\u838e\u62c9", "\u7b39\u91ce", "\u806a\u7f8e", "\u806a", "\u5c0f\u767e\u5408", "\u6563\u5175", "\u5bb3\u6015\u7684\u5c0f\u5218", "\u8212\u4f2f\u7279", "\u8212\u8328", "\u6d77\u9f99", "\u4e16\u5b50", "\u8c22\u5c14\u76d6", "\u5bb6\u4e01", "\u5546\u534e", "\u6c99\u5bc5", "\u963f\u5347", "\u67f4\u7530", "\u963f\u8302", "\u5f0f\u5927\u5c06", "\u6e05\u6c34", "\u5fd7\u6751\u52d8\u5175\u536b", "\u65b0\u4e4b\u4e1e", "\u5fd7\u7ec7", "\u77f3\u5934", "\u8bd7\u7fbd", "\u8bd7\u7b60", "\u77f3\u58ee", "\u7fd4\u592a", "\u6b63\u4e8c", "\u5468\u5e73", "\u8212\u6768", "\u9f50\u683c\u8299\u4e3d\u96c5", "\u5973\u58eb", "\u601d\u52e4", "\u516d\u6307\u4e54\u745f", "\u611a\u4eba\u4f17\u5c0f\u5175d", "\u611a\u4eba\u4f17\u5c0f\u5175a", "\u611a\u4eba\u4f17\u5c0f\u5175b", "\u611a\u4eba\u4f17\u5c0f\u5175c", "\u5434\u8001\u4e94", "\u5434\u8001\u4e8c", "\u6ed1\u5934\u9b3c", "\u8a00\u7b11", "\u5434\u8001\u4e03", "\u58eb\u5175h", "\u58eb\u5175i", "\u58eb\u5175a", "\u58eb\u5175b", "\u58eb\u5175c", "\u58eb\u5175d", "\u58eb\u5175e", "\u58eb\u5175f", "\u58eb\u5175g", "\u594f\u592a", "\u65af\u5766\u5229", "\u6387\u661f\u652b\u8fb0\u5929\u541b", "\u5c0f\u5934", "\u5927\u6b66", "\u9676\u4e49\u9686", "\u6749\u672c", "\u82cf\u897f", "\u5acc\u7591\u4ebaa", "\u5acc\u7591\u4ebab", "\u5acc\u7591\u4ebac", "\u5acc\u7591\u4ebad", "\u65af\u4e07", "\u5251\u5ba2a", "\u5251\u5ba2b", "\u963f\u4e8c", "\u5fe0\u80dc", "\u5fe0\u592b", "\u963f\u656c", "\u5b5d\u5229", "\u9e70\u53f8\u8fdb", "\u9ad8\u5c71", "\u4e5d\u6761\u5b5d\u884c", "\u6bc5", "\u7af9\u5185", "\u62d3\u771f", "\u5353\u4e5f", "\u592a\u90ce\u4e38", "\u6cf0\u52d2", "\u624b\u5c9b", "\u54f2\u5e73", "\u54f2\u592b", "\u6258\u514b", "\u5927boss", "\u963f\u5f3a", "\u6258\u5c14\u5fb7\u62c9", "\u65c1\u89c2\u8005", "\u5929\u6210", "\u963f\u5927", "\u8482\u739b\u4e4c\u65af", "\u63d0\u7c73", "\u6237\u7530", "\u963f\u4e09", "\u4e00\u8d77\u7684\u4eba", "\u5fb7\u7530", "\u5fb7\u957f", "\u667a\u6811", "\u5229\u5f66", "\u80d6\u4e4e\u4e4e\u7684\u65c5\u884c\u8005", "\u85cf\u5b9d\u4ebaa", "\u85cf\u5b9d\u4ebab", "\u85cf\u5b9d\u4ebac", "\u85cf\u5b9d\u4ebad", "\u963f\u7947", "\u6052\u96c4", "\u9732\u5b50", "\u8bdd\u5267\u56e2\u56e2\u957f", "\u5185\u6751", "\u4e0a\u91ce", "\u4e0a\u6749", "\u8001\u6234", "\u8001\u9ad8", "\u8001\u8d3e", "\u8001\u58a8", "\u8001\u5b59", "\u5929\u67a2\u661f", "\u8001\u4e91", "\u6709\u4e50\u658b", "\u4e11\u96c4", "\u4e4c\u7ef4", "\u74e6\u4eac", "\u83f2\u5c14\u6208\u9edb\u7279", "\u7ef4\u591a\u5229\u4e9a", "\u8587\u5c14", "\u74e6\u683c\u7eb3", "\u963f\u5916", "\u4f8d\u5973", "\u74e6\u62c9", "\u671b\u96c5", "\u5b9b\u70df", "\u742c\u7389", "\u6218\u58eba", "\u6218\u58ebb", "\u6e21\u8fba", "\u6e21\u90e8", "\u963f\u4f1f", "\u6587\u749f", "\u6587\u6e0a", "\u97e6\u5c14\u7eb3", "\u738b\u6273\u624b", "\u6b66\u6c9b", "\u6653\u98de", "\u8f9b\u7a0b", "\u661f\u706b", "\u661f\u7a00", "\u8f9b\u79c0", "\u79c0\u534e", "\u963f\u65ed", "\u5f90\u5218\u5e08", "\u77e2\u90e8", "\u516b\u6728", "\u5c71\u4e0a", "\u963f\u9633", "\u989c\u7b11", "\u5eb7\u660e", "\u6cf0\u4e45", "\u5b89\u6b66", "\u77e2\u7530\u5e78\u559c", "\u77e2\u7530\u8f9b\u559c", "\u4e49\u575a", "\u83ba\u513f", "\u76c8\u4e30", "\u5b9c\u5e74", "\u94f6\u674f", "\u9038\u8f69", "\u6a2a\u5c71", "\u6c38\u8d35", "\u6c38\u4e1a", "\u5609\u4e45", "\u5409\u5ddd", "\u4e49\u9ad8", "\u7528\u9ad8", "\u9633\u592a", "\u5143\u84c9", "\u73a5\u8f89", "\u6bd3\u534e", "\u6709\u9999", "\u5e78\u4e5f", "\u7531\u771f", "\u7ed3\u83dc", "\u97f5\u5b81", "\u767e\u5408", "\u767e\u5408\u534e", "\u5c24\u82cf\u6ce2\u592b", "\u88d5\u5b50", "\u60a0\u7b56", "\u60a0\u4e5f", "\u4e8e\u5ae3", "\u67da\u5b50", "\u8001\u90d1", "\u6b63\u8302", "\u5fd7\u6210", "\u82b7\u5de7", "\u77e5\u6613", "\u652f\u652f", "\u5468\u826f", "\u73e0\u51fd", "\u795d\u660e", "\u795d\u6d9b"],
54
+ "symbols": ["_", ",", ".", "!", "?", "-", "~", "\u2026", "A", "E", "I", "N", "O", "Q", "U", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "r", "s", "t", "u", "v", "w", "y", "z", "\u0283", "\u02a7", "\u02a6", "\u026f", "\u0279", "\u0259", "\u0265", "\u207c", "\u02b0", "`", "\u2192", "\u2193", "\u2191", " "]
55
+ }
README.md CHANGED
@@ -1,11 +1,231 @@
1
- ---
2
- title: MoeGoe Simple API
3
- emoji: 🌍
4
- colorFrom: yellow
5
- colorTo: purple
6
- sdk: docker
7
- pinned: false
8
- license: mit
9
- ---
10
-
11
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # MoeGoe-Simple-API
2
+
3
+
4
+
5
+ MoeGoe-Simple-API 是一个易部署的api,可以通过api的方式调用语音合成,可用于聊天机器人等,目前支持的功能有语音合成和语音转换。
6
+
7
+ - 可导入模型
8
+ - 支持加载多模型,可以将多个模型合并为一个新的id对应角色模型的映射表
9
+ <details><summary>点击预览返回的映射表</summary><pre><code>
10
+ [
11
+ {
12
+ "0": "綾地寧々"
13
+ },
14
+ {
15
+ "1": "在原七海"
16
+ },
17
+ {
18
+ "2": "小茸"
19
+ },
20
+ {
21
+ "3": "唐乐吟"
22
+ },
23
+ {
24
+ "...": "..."
25
+ },
26
+ {
27
+ "196": "ルイズ"
28
+ },
29
+ {
30
+ "197": "ティファニア"
31
+ }
32
+ ]
33
+ </code></pre></details>
34
+
35
+ # 如何部署
36
+
37
+ ## Docker部署
38
+
39
+ ### docker镜像拉取脚本
40
+
41
+ ```
42
+ bash -c "$(wget -O- https://gist.githubusercontent.com/Artrajz/b2c02499f91c3a51b8b48f1a3c9a7ead/raw/e3033f1b222868b4b0f1b522e52e18217460ff91/moegoe-simple-api-installer-latest.sh)"
43
+ ```
44
+
45
+ 镜像大小为5g,所以拉取会比较慢,拉取成功后由于没有导入vits模型所以无法使用,需要按以下步骤导入模型
46
+
47
+ ### 下载VITS模型
48
+
49
+ VITS模型放入`/usr/local/moegoe-simple-api/Model`文件夹中,模型文件夹中要有.pth和config.json文件
50
+
51
+ <details><summary>点击查看模型目录结构</summary><pre><code>
52
+ ├─Model
53
+ │ ├─g
54
+ │ │ config.json
55
+ │ │ G_953000.pth
56
+ │ │
57
+ │ ├─Nene_Nanami_Rong_Tang
58
+ │ │ 1374_epochs.pth
59
+ │ │ config.json
60
+ │ │
61
+ │ └─Zero_no_tsukaima
62
+ │ 1158_epochs.pth
63
+ │ config.json
64
+ </code></pre></details>
65
+
66
+ ### 修改模型路径
67
+
68
+ 在 `/usr/local/moegoe-simple-api/config.py` 中修改模型路径
69
+
70
+ <details><summary>点击查看config.py模型路径填写示例</summary><pre><code>
71
+ vits模型路径填写方法,MODEL_LIST中的每一行是
72
+ [ABS_PATH+"/Model/{模型文件夹}/{.pth模型}", ABS_PATH+"/Model/{模型文件夹}/config.json"],
73
+ 也可以写相对路径或绝对路径,由于windows和linux路径写法不同,用上面的写法或绝对路径最稳妥
74
+ 示例:
75
+ MODEL_LIST = [
76
+ [ABS_PATH+"/Model/Nene_Nanami_Rong_Tang/1374_epochs.pth", ABS_PATH+"/Model/Nene_Nanami_Rong_Tang/config.json"],
77
+ [ABS_PATH+"/Model/Zero_no_tsukaima/1158_epochs.pth", ABS_PATH+"/Model/Zero_no_tsukaima/config.json"],
78
+ [ABS_PATH+"/Model/g/G_953000.pth", ABS_PATH+"/Model/g/config.json"],
79
+ ]
80
+ </code></pre></details>
81
+
82
+
83
+ ### 开始使用!
84
+
85
+ 终端输入`docker compose up -d`
86
+
87
+ 或再次执行拉取脚本
88
+
89
+ ### 镜像更新方法
90
+
91
+ 重新执行docker镜像拉取脚本即可
92
+
93
+
94
+ ## 直接部署
95
+
96
+ 1. 下载VITS模型并放入`Model`文件夹中
97
+ 2. 在 `config.py` 中修改模型路径
98
+ 3. 安装python依赖(建议用conda虚拟环境) `pip install -r requirements.txt`
99
+ 4. 开始使用!`python app.py`
100
+
101
+ # 参数
102
+
103
+ ## 语音合成voice
104
+
105
+ | Name | Parameter | Is must | Default | Value | Instruction |
106
+ | ------------- | --------- | ------- | ------- | ------------ | ------------------------------------------------ |
107
+ | 合成文本 | text | true | | text | |
108
+ | 角色id | id | false | 0 | (number) | |
109
+ | 音频格式 | format | false | wav | wav,ogg,silk | silk支持tx系语音 |
110
+ | 文本语言 | lang | false | mix | zh,ja,mix | 当lang=mix时,文本应该用[ZH] 或 [JA] 包裹 |
111
+ | 语音长度/语速 | speed | false | 1.0 | (number) | 调节语音长度,相当于调节语速,该数值越大语速越慢 |
112
+
113
+ ## 语音转换voice conversion
114
+
115
+ | Name | Parameter | Is must | Default | Value | Instruction |
116
+ | ---------- | ----------- | ------- | ------- | ---------- | ---------------------- |
117
+ | 上传音频 | upload | true | | audio file | 只支持wav和ogg |
118
+ | 源角色id | original_id | true | | (number) | 上传文件所使用的角色id |
119
+ | 目标角色id | target_id | true | | (number) | 要转换的目标角色id |
120
+
121
+ # 调用方法
122
+
123
+ ## GET
124
+
125
+ - GET/POST http://127.0.0.1:23456/voice/speakers
126
+
127
+ 返回id对应角色的映射表(json格式)
128
+
129
+ - GET http://127.0.0.1/voice?text=[JA]text[JA][ZH]text[ZH]&id=0&format=wav&lang=mix
130
+
131
+ 返回wav音频文件
132
+
133
+ - GET http://127.0.0.1/voice?text=[JA]text[JA][ZH]text[ZH]&id=0&format=ogg&lang=mix
134
+
135
+ 返回ogg音频文件
136
+
137
+ - GET http://127.0.0.1/voice?text=text&lang=zh
138
+
139
+ 设定语言为zh,则文本无需[ZH]包裹
140
+
141
+ - GET http://127.0.0.1/voice?text=text&lang=ja
142
+
143
+ 设定语言为ja,则文本无需[JA]包裹
144
+
145
+ - GET http://127.0.0.1/voice?text=text&id=142&format=wav&lang=zh&speed=1.4
146
+
147
+ 文本为text,角色id为142,音频格式为wav,文本语言为zh,语速为1.4
148
+
149
+ ## POST
150
+
151
+ - python
152
+
153
+ ```python
154
+ import re
155
+ import requests
156
+ import json
157
+ import os
158
+ import random
159
+ import string
160
+ from requests_toolbelt.multipart.encoder import MultipartEncoder
161
+
162
+ abs_path = os.path.dirname(__file__)
163
+
164
+ def voice_speakers():
165
+ url = "http://127.0.0.1:23456/voice/speakers"
166
+
167
+ res = requests.post(url=url)
168
+ json = res.json()
169
+ for i in json:
170
+ print(i)
171
+
172
+ def voice():
173
+ post_json = json.dumps({
174
+ "text":"需要合成的文本",
175
+ "id":142,
176
+ "format":"wav",
177
+ "lang":"zh",
178
+ "speed":1.4,
179
+ })
180
+ headers={'content-type':'application/json'}
181
+ url = "http://127.0.0.1:23456/voice"
182
+
183
+ res = requests.post(url=url,data=post_json,headers=headers)
184
+ fname = re.findall("filename=(.+)", res.headers["Content-Disposition"])[0]
185
+ path = f"{abs_path}/{fname}"
186
+ with open(path, "wb") as f:
187
+ f.write(res.content)
188
+
189
+ def voice_conversion(upload_name):
190
+ upload_path = f'{abs_path}/{upload_name}'
191
+ upload_type = f'audio/{upload_name.split(".")[1]}' #wav,ogg
192
+
193
+ fields = {
194
+ "upload": (upload_name, open(upload_path,'rb'),upload_type),
195
+ "original_id": "142",
196
+ "target_id": "92",
197
+ }
198
+ boundary = '----VoiceConversionFormBoundary' \
199
+ + ''.join(random.sample(string.ascii_letters + string.digits, 16))
200
+ m = MultipartEncoder(fields=fields, boundary=boundary)
201
+
202
+ headers = {"Content-Type": m.content_type}
203
+ url = "http://127.0.0.1:23456/voice/conversion"
204
+
205
+ res = requests.post(url=url,data=m,headers=headers)
206
+ fname = re.findall("filename=(.+)", res.headers["Content-Disposition"])[0]
207
+ path = f"{abs_path}/{fname}"
208
+ with open(path, "wb") as f:
209
+ f.write(res.content)
210
+ ```
211
+
212
+ # 可能遇到的问题
213
+
214
+ ~~本人遇到过的问题~~
215
+
216
+ ### 运行后服务器无响应
217
+
218
+ 可能是内存不足,可以尝试减少模型加载数量。
219
+
220
+ ### 模型推理时服务器无响应
221
+
222
+ 可能是同时处理多个推理任务导致CPU堵塞,可以尝试在*voice.py*中取消以下**两行**代码的注释,意思是让pytorch只使用1个物理CPU核心,防止一个任务抢占过多CPU资源。
223
+
224
+ ```python
225
+ import torch
226
+ torch.set_num_threads(1)
227
+ ```
228
+
229
+ # 鸣谢
230
+
231
+ 该项目基于[CjangCjengh](https://github.com/CjangCjengh)的[MoeGoe](https://github.com/CjangCjengh/MoeGoe)
app.py ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ import logging
4
+ import uuid
5
+
6
+ from flask import Flask, request, send_file, jsonify
7
+ from werkzeug.utils import secure_filename
8
+ from flask_apscheduler import APScheduler
9
+
10
+ from utils import clean_folder, merge_model
11
+
12
+ app = Flask(__name__)
13
+ app.config.from_pyfile("config.py")
14
+
15
+ scheduler = APScheduler()
16
+ scheduler.init_app(app)
17
+ scheduler.start()
18
+
19
+ logging.getLogger('numba').setLevel(logging.WARNING)
20
+
21
+ voice_obj, voice_speakers = merge_model(app.config["MODEL_LIST"])
22
+
23
+ if not os.path.exists(app.config['UPLOAD_FOLDER']):
24
+ try:
25
+ os.mkdir(app.config['UPLOAD_FOLDER'])
26
+ except:
27
+ pass
28
+
29
+
30
+ @app.route('/')
31
+ @app.route('/voice/')
32
+ def index():
33
+ return "usage:https://github.com/Artrajz/MoeGoe-Simple-API#readme"
34
+
35
+
36
+ @app.route('/voice/speakers', methods=["GET", "POST"])
37
+ def voice_speakers_api():
38
+ speakers_list = voice_speakers
39
+ return jsonify(speakers_list)
40
+
41
+
42
+ @app.route('/voice', methods=["GET", "POST"])
43
+ def voice_api():
44
+ if request.method == "GET":
45
+ text = request.args.get("text")
46
+ speaker_id = int(request.args.get("id", 0))
47
+ format = request.args.get("format", "wav")
48
+ lang = request.args.get("lang", "mix")
49
+ speed = float(request.args.get("speed", 1.0))
50
+ elif request.method == "POST":
51
+ json_data = request.json
52
+ text = json_data["text"]
53
+ speaker_id = int(json_data["id"])
54
+ format = json_data["format"]
55
+ lang = json_data["lang"]
56
+ speed = float(json_data["speed"])
57
+
58
+ if lang.upper() == "ZH":
59
+ text = f"[ZH]{text}[ZH]"
60
+ elif lang.upper() == "JA":
61
+ text = f"[JA]{text}[JA]"
62
+
63
+ real_id = voice_obj[speaker_id][0]
64
+ real_obj = voice_obj[speaker_id][1]
65
+
66
+ output, file_type, fname = real_obj.generate(text, real_id, format, speed)
67
+
68
+ return send_file(path_or_file=output, mimetype=file_type, download_name=fname)
69
+
70
+
71
+ @app.route('/voice/conversion', methods=["GET", "POST"])
72
+ def voice_conversion_api():
73
+ if request.method == "GET":
74
+ return jsonify("method should be POST")
75
+ if request.method == "POST":
76
+ # json_data = request.json
77
+ voice = request.files['upload']
78
+ original_id = int(request.form["original_id"])
79
+ target_id = int(request.form["target_id"])
80
+
81
+ form = {}
82
+
83
+ format = voice.filename.split(".")[1]
84
+
85
+ fname = secure_filename(str(uuid.uuid1()) + "." + voice.filename.split(".")[1])
86
+ voice.save(os.path.join(app.config['UPLOAD_FOLDER'], fname))
87
+
88
+ real_original_id = int(voice_obj[original_id][0])
89
+ real_target_id = int(voice_obj[target_id][0])
90
+ real_obj = voice_obj[original_id][1]
91
+ real_target_obj = voice_obj[target_id][1]
92
+
93
+ if voice_obj[original_id][2] != voice_obj[target_id][2]:
94
+ form["status"] = "error"
95
+ form["message"] = "speaker IDs are in diffrent Model!"
96
+ return form
97
+
98
+ output = real_obj.voice_conversion(os.path.join(app.config['UPLOAD_FOLDER'], fname),
99
+ real_original_id, real_target_id, format)
100
+ file_type = f"audio/{format}"
101
+
102
+ return send_file(path_or_file=output, mimetype=file_type, download_name=fname)
103
+ # return output
104
+
105
+
106
+ # 定时清理临时文件,每小时清一次
107
+ @scheduler.task('interval', id='随便写', seconds=3600, misfire_grace_time=900)
108
+ def clean_task():
109
+ clean_folder(app.config["UPLOAD_FOLDER"])
110
+ clean_folder(app.config["SILK_OUT_PATH"])
111
+
112
+
113
+ if __name__ == '__main__':
114
+ app.run(host='0.0.0.0', port=app.config["PORT"]) # 如果对外开放用这个,docker部署也用这个
115
+ # app.run(host='127.0.0.1', port=app.config["PORT"], debug=True) # 本地运行、调试
attentions.py ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ import commons
7
+ from modules import LayerNorm
8
+
9
+
10
+ class Encoder(nn.Module):
11
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
12
+ super().__init__()
13
+ self.hidden_channels = hidden_channels
14
+ self.filter_channels = filter_channels
15
+ self.n_heads = n_heads
16
+ self.n_layers = n_layers
17
+ self.kernel_size = kernel_size
18
+ self.p_dropout = p_dropout
19
+ self.window_size = window_size
20
+
21
+ self.drop = nn.Dropout(p_dropout)
22
+ self.attn_layers = nn.ModuleList()
23
+ self.norm_layers_1 = nn.ModuleList()
24
+ self.ffn_layers = nn.ModuleList()
25
+ self.norm_layers_2 = nn.ModuleList()
26
+ for i in range(self.n_layers):
27
+ self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
28
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
29
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
30
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
31
+
32
+ def forward(self, x, x_mask):
33
+ attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
34
+ x = x * x_mask
35
+ for i in range(self.n_layers):
36
+ y = self.attn_layers[i](x, x, attn_mask)
37
+ y = self.drop(y)
38
+ x = self.norm_layers_1[i](x + y)
39
+
40
+ y = self.ffn_layers[i](x, x_mask)
41
+ y = self.drop(y)
42
+ x = self.norm_layers_2[i](x + y)
43
+ x = x * x_mask
44
+ return x
45
+
46
+
47
+ class Decoder(nn.Module):
48
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
49
+ super().__init__()
50
+ self.hidden_channels = hidden_channels
51
+ self.filter_channels = filter_channels
52
+ self.n_heads = n_heads
53
+ self.n_layers = n_layers
54
+ self.kernel_size = kernel_size
55
+ self.p_dropout = p_dropout
56
+ self.proximal_bias = proximal_bias
57
+ self.proximal_init = proximal_init
58
+
59
+ self.drop = nn.Dropout(p_dropout)
60
+ self.self_attn_layers = nn.ModuleList()
61
+ self.norm_layers_0 = nn.ModuleList()
62
+ self.encdec_attn_layers = nn.ModuleList()
63
+ self.norm_layers_1 = nn.ModuleList()
64
+ self.ffn_layers = nn.ModuleList()
65
+ self.norm_layers_2 = nn.ModuleList()
66
+ for i in range(self.n_layers):
67
+ self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
68
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
69
+ self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
70
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
71
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
72
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
73
+
74
+ def forward(self, x, x_mask, h, h_mask):
75
+ """
76
+ x: decoder input
77
+ h: encoder output
78
+ """
79
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
80
+ encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
81
+ x = x * x_mask
82
+ for i in range(self.n_layers):
83
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
84
+ y = self.drop(y)
85
+ x = self.norm_layers_0[i](x + y)
86
+
87
+ y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
88
+ y = self.drop(y)
89
+ x = self.norm_layers_1[i](x + y)
90
+
91
+ y = self.ffn_layers[i](x, x_mask)
92
+ y = self.drop(y)
93
+ x = self.norm_layers_2[i](x + y)
94
+ x = x * x_mask
95
+ return x
96
+
97
+
98
+ class MultiHeadAttention(nn.Module):
99
+ def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
100
+ super().__init__()
101
+ assert channels % n_heads == 0
102
+
103
+ self.channels = channels
104
+ self.out_channels = out_channels
105
+ self.n_heads = n_heads
106
+ self.p_dropout = p_dropout
107
+ self.window_size = window_size
108
+ self.heads_share = heads_share
109
+ self.block_length = block_length
110
+ self.proximal_bias = proximal_bias
111
+ self.proximal_init = proximal_init
112
+ self.attn = None
113
+
114
+ self.k_channels = channels // n_heads
115
+ self.conv_q = nn.Conv1d(channels, channels, 1)
116
+ self.conv_k = nn.Conv1d(channels, channels, 1)
117
+ self.conv_v = nn.Conv1d(channels, channels, 1)
118
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
119
+ self.drop = nn.Dropout(p_dropout)
120
+
121
+ if window_size is not None:
122
+ n_heads_rel = 1 if heads_share else n_heads
123
+ rel_stddev = self.k_channels**-0.5
124
+ self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
125
+ self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
126
+
127
+ nn.init.xavier_uniform_(self.conv_q.weight)
128
+ nn.init.xavier_uniform_(self.conv_k.weight)
129
+ nn.init.xavier_uniform_(self.conv_v.weight)
130
+ if proximal_init:
131
+ with torch.no_grad():
132
+ self.conv_k.weight.copy_(self.conv_q.weight)
133
+ self.conv_k.bias.copy_(self.conv_q.bias)
134
+
135
+ def forward(self, x, c, attn_mask=None):
136
+ q = self.conv_q(x)
137
+ k = self.conv_k(c)
138
+ v = self.conv_v(c)
139
+
140
+ x, self.attn = self.attention(q, k, v, mask=attn_mask)
141
+
142
+ x = self.conv_o(x)
143
+ return x
144
+
145
+ def attention(self, query, key, value, mask=None):
146
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
147
+ b, d, t_s, t_t = (*key.size(), query.size(2))
148
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
149
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
150
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
151
+
152
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
153
+ if self.window_size is not None:
154
+ assert t_s == t_t, "Relative attention is only available for self-attention."
155
+ key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
156
+ rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
157
+ scores_local = self._relative_position_to_absolute_position(rel_logits)
158
+ scores = scores + scores_local
159
+ if self.proximal_bias:
160
+ assert t_s == t_t, "Proximal bias is only available for self-attention."
161
+ scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
162
+ if mask is not None:
163
+ scores = scores.masked_fill(mask == 0, -1e4)
164
+ if self.block_length is not None:
165
+ assert t_s == t_t, "Local attention is only available for self-attention."
166
+ block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
167
+ scores = scores.masked_fill(block_mask == 0, -1e4)
168
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
169
+ p_attn = self.drop(p_attn)
170
+ output = torch.matmul(p_attn, value)
171
+ if self.window_size is not None:
172
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
173
+ value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
174
+ output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
175
+ output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
176
+ return output, p_attn
177
+
178
+ def _matmul_with_relative_values(self, x, y):
179
+ """
180
+ x: [b, h, l, m]
181
+ y: [h or 1, m, d]
182
+ ret: [b, h, l, d]
183
+ """
184
+ ret = torch.matmul(x, y.unsqueeze(0))
185
+ return ret
186
+
187
+ def _matmul_with_relative_keys(self, x, y):
188
+ """
189
+ x: [b, h, l, d]
190
+ y: [h or 1, m, d]
191
+ ret: [b, h, l, m]
192
+ """
193
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
194
+ return ret
195
+
196
+ def _get_relative_embeddings(self, relative_embeddings, length):
197
+ max_relative_position = 2 * self.window_size + 1
198
+ # Pad first before slice to avoid using cond ops.
199
+ pad_length = max(length - (self.window_size + 1), 0)
200
+ slice_start_position = max((self.window_size + 1) - length, 0)
201
+ slice_end_position = slice_start_position + 2 * length - 1
202
+ if pad_length > 0:
203
+ padded_relative_embeddings = F.pad(
204
+ relative_embeddings,
205
+ commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
206
+ else:
207
+ padded_relative_embeddings = relative_embeddings
208
+ used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
209
+ return used_relative_embeddings
210
+
211
+ def _relative_position_to_absolute_position(self, x):
212
+ """
213
+ x: [b, h, l, 2*l-1]
214
+ ret: [b, h, l, l]
215
+ """
216
+ batch, heads, length, _ = x.size()
217
+ # Concat columns of pad to shift from relative to absolute indexing.
218
+ x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
219
+
220
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
221
+ x_flat = x.view([batch, heads, length * 2 * length])
222
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
223
+
224
+ # Reshape and slice out the padded elements.
225
+ x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
226
+ return x_final
227
+
228
+ def _absolute_position_to_relative_position(self, x):
229
+ """
230
+ x: [b, h, l, l]
231
+ ret: [b, h, l, 2*l-1]
232
+ """
233
+ batch, heads, length, _ = x.size()
234
+ # padd along column
235
+ x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
236
+ x_flat = x.view([batch, heads, length**2 + length*(length -1)])
237
+ # add 0's in the beginning that will skew the elements after reshape
238
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
239
+ x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
240
+ return x_final
241
+
242
+ def _attention_bias_proximal(self, length):
243
+ """Bias for self-attention to encourage attention to close positions.
244
+ Args:
245
+ length: an integer scalar.
246
+ Returns:
247
+ a Tensor with shape [1, 1, length, length]
248
+ """
249
+ r = torch.arange(length, dtype=torch.float32)
250
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
251
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
252
+
253
+
254
+ class FFN(nn.Module):
255
+ def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
256
+ super().__init__()
257
+ self.in_channels = in_channels
258
+ self.out_channels = out_channels
259
+ self.filter_channels = filter_channels
260
+ self.kernel_size = kernel_size
261
+ self.p_dropout = p_dropout
262
+ self.activation = activation
263
+ self.causal = causal
264
+
265
+ if causal:
266
+ self.padding = self._causal_padding
267
+ else:
268
+ self.padding = self._same_padding
269
+
270
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
271
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
272
+ self.drop = nn.Dropout(p_dropout)
273
+
274
+ def forward(self, x, x_mask):
275
+ x = self.conv_1(self.padding(x * x_mask))
276
+ if self.activation == "gelu":
277
+ x = x * torch.sigmoid(1.702 * x)
278
+ else:
279
+ x = torch.relu(x)
280
+ x = self.drop(x)
281
+ x = self.conv_2(self.padding(x * x_mask))
282
+ return x * x_mask
283
+
284
+ def _causal_padding(self, x):
285
+ if self.kernel_size == 1:
286
+ return x
287
+ pad_l = self.kernel_size - 1
288
+ pad_r = 0
289
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
290
+ x = F.pad(x, commons.convert_pad_shape(padding))
291
+ return x
292
+
293
+ def _same_padding(self, x):
294
+ if self.kernel_size == 1:
295
+ return x
296
+ pad_l = (self.kernel_size - 1) // 2
297
+ pad_r = self.kernel_size // 2
298
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
299
+ x = F.pad(x, commons.convert_pad_shape(padding))
300
+ return x
commons.py ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+ import torch.jit
4
+
5
+
6
+ def script_method(fn, _rcb=None):
7
+ return fn
8
+
9
+
10
+ def script(obj, optimize=True, _frames_up=0, _rcb=None):
11
+ return obj
12
+
13
+
14
+ torch.jit.script_method = script_method
15
+ torch.jit.script = script
16
+
17
+
18
+ def init_weights(m, mean=0.0, std=0.01):
19
+ classname = m.__class__.__name__
20
+ if classname.find("Conv") != -1:
21
+ m.weight.data.normal_(mean, std)
22
+
23
+
24
+ def get_padding(kernel_size, dilation=1):
25
+ return int((kernel_size*dilation - dilation)/2)
26
+
27
+
28
+ def intersperse(lst, item):
29
+ result = [item] * (len(lst) * 2 + 1)
30
+ result[1::2] = lst
31
+ return result
32
+
33
+
34
+ def slice_segments(x, ids_str, segment_size=4):
35
+ ret = torch.zeros_like(x[:, :, :segment_size])
36
+ for i in range(x.size(0)):
37
+ idx_str = ids_str[i]
38
+ idx_end = idx_str + segment_size
39
+ ret[i] = x[i, :, idx_str:idx_end]
40
+ return ret
41
+
42
+
43
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
44
+ b, d, t = x.size()
45
+ if x_lengths is None:
46
+ x_lengths = t
47
+ ids_str_max = x_lengths - segment_size + 1
48
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
49
+ ret = slice_segments(x, ids_str, segment_size)
50
+ return ret, ids_str
51
+
52
+
53
+ def subsequent_mask(length):
54
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
55
+ return mask
56
+
57
+
58
+ @torch.jit.script
59
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
60
+ n_channels_int = n_channels[0]
61
+ in_act = input_a + input_b
62
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
63
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
64
+ acts = t_act * s_act
65
+ return acts
66
+
67
+
68
+ def convert_pad_shape(pad_shape):
69
+ l = pad_shape[::-1]
70
+ pad_shape = [item for sublist in l for item in sublist]
71
+ return pad_shape
72
+
73
+
74
+ def sequence_mask(length, max_length=None):
75
+ if max_length is None:
76
+ max_length = length.max()
77
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
78
+ return x.unsqueeze(0) < length.unsqueeze(1)
79
+
80
+
81
+ def generate_path(duration, mask):
82
+ """
83
+ duration: [b, 1, t_x]
84
+ mask: [b, 1, t_y, t_x]
85
+ """
86
+ device = duration.device
87
+
88
+ b, _, t_y, t_x = mask.shape
89
+ cum_duration = torch.cumsum(duration, -1)
90
+
91
+ cum_duration_flat = cum_duration.view(b * t_x)
92
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
93
+ path = path.view(b, t_x, t_y)
94
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
95
+ path = path.unsqueeze(1).transpose(2,3) * mask
96
+ return path
config.py ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import sys
3
+
4
+ JSON_AS_ASCII = False
5
+ MAX_CONTENT_LENGTH = 5242880
6
+
7
+ # 端口
8
+ PORT = 7860
9
+ # 项目的绝对路径
10
+ ABS_PATH = os.path.join(os.path.dirname(os.path.realpath(sys.argv[0])))
11
+ # 上传文件的临时路径,非必要不要动
12
+ UPLOAD_FOLDER = ABS_PATH + "/upload"
13
+ # silk文件输出的临时路径,非必要不要动
14
+ SILK_OUT_PATH = ABS_PATH + "/out_silk"
15
+
16
+ '''
17
+ vits模型路径填写方法,MODEL_LIST中的每一行是
18
+ [ABS_PATH+"/Model/{模型文件夹}/{.pth模型}", ABS_PATH+"/Model/{模型文件夹}/config.json"],
19
+ 也可以写相对路径或绝对路径,由于windows和linux路径写法不同,用上面的写法或绝对路径最稳妥
20
+ 示例:
21
+ MODEL_LIST = [
22
+ [ABS_PATH+"/Model/Nene_Nanami_Rong_Tang/1374_epochs.pth", ABS_PATH+"/Model/Nene_Nanami_Rong_Tang/config.json"],
23
+ [ABS_PATH+"/Model/Zero_no_tsukaima/1158_epochs.pth", ABS_PATH+"/Model/Zero_no_tsukaima/config.json"],
24
+ [ABS_PATH+"/Model/g/G_953000.pth", ABS_PATH+"/Model/g/config.json"],
25
+ ]
26
+ '''
27
+ # 模型加载列表
28
+ MODEL_LIST = [
29
+ [ABS_PATH+"/Model/Nene_Nanami_Rong_Tang/1374_epochs.pth", ABS_PATH+"/Model/Nene_Nanami_Rong_Tang/config.json"],
30
+ [ABS_PATH+"/Model/Zero_no_tsukaima/1158_epochs.pth", ABS_PATH+"/Model/Zero_no_tsukaima/config.json"],
31
+ [ABS_PATH+"/Model/g/G_953000.pth", ABS_PATH+"/Model/g/config.json"],
32
+ ]
docker-compose.yaml ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ version: '3.4'
2
+ services:
3
+ moegoe:
4
+ image: artrajz/moegoe-simple-api:latest
5
+ restart: always
6
+ ports:
7
+ - 23456:23456
8
+ environment:
9
+ LANG: 'C.UTF-8'
10
+ volumes:
11
+ - ./Model:/app/Model # 挂载模型文件夹
12
+ - ./config.py:/app/config.py # 挂载配置文件
hubert_model.py ADDED
@@ -0,0 +1,221 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ from typing import Optional, Tuple
3
+ import random
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ import torch.nn.functional as F
8
+ from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
9
+
10
+ class Hubert(nn.Module):
11
+ def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
12
+ super().__init__()
13
+ self._mask = mask
14
+ self.feature_extractor = FeatureExtractor()
15
+ self.feature_projection = FeatureProjection()
16
+ self.positional_embedding = PositionalConvEmbedding()
17
+ self.norm = nn.LayerNorm(768)
18
+ self.dropout = nn.Dropout(0.1)
19
+ self.encoder = TransformerEncoder(
20
+ nn.TransformerEncoderLayer(
21
+ 768, 12, 3072, activation="gelu", batch_first=True
22
+ ),
23
+ 12,
24
+ )
25
+ self.proj = nn.Linear(768, 256)
26
+
27
+ self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
28
+ self.label_embedding = nn.Embedding(num_label_embeddings, 256)
29
+
30
+ def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
31
+ mask = None
32
+ if self.training and self._mask:
33
+ mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
34
+ x[mask] = self.masked_spec_embed.to(x.dtype)
35
+ return x, mask
36
+
37
+ def encode(
38
+ self, x: torch.Tensor, layer: Optional[int] = None
39
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
40
+ x = self.feature_extractor(x)
41
+ x = self.feature_projection(x.transpose(1, 2))
42
+ x, mask = self.mask(x)
43
+ x = x + self.positional_embedding(x)
44
+ x = self.dropout(self.norm(x))
45
+ x = self.encoder(x, output_layer=layer)
46
+ return x, mask
47
+
48
+ def logits(self, x: torch.Tensor) -> torch.Tensor:
49
+ logits = torch.cosine_similarity(
50
+ x.unsqueeze(2),
51
+ self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
52
+ dim=-1,
53
+ )
54
+ return logits / 0.1
55
+
56
+ def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
57
+ x, mask = self.encode(x)
58
+ x = self.proj(x)
59
+ logits = self.logits(x)
60
+ return logits, mask
61
+
62
+
63
+ class HubertSoft(Hubert):
64
+ def __init__(self):
65
+ super().__init__()
66
+
67
+ @torch.inference_mode()
68
+ def units(self, wav: torch.Tensor) -> torch.Tensor:
69
+ wav = F.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
70
+ x, _ = self.encode(wav)
71
+ return self.proj(x)
72
+
73
+
74
+ class FeatureExtractor(nn.Module):
75
+ def __init__(self):
76
+ super().__init__()
77
+ self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
78
+ self.norm0 = nn.GroupNorm(512, 512)
79
+ self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
80
+ self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
81
+ self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
82
+ self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
83
+ self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
84
+ self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
85
+
86
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
87
+ x = F.gelu(self.norm0(self.conv0(x)))
88
+ x = F.gelu(self.conv1(x))
89
+ x = F.gelu(self.conv2(x))
90
+ x = F.gelu(self.conv3(x))
91
+ x = F.gelu(self.conv4(x))
92
+ x = F.gelu(self.conv5(x))
93
+ x = F.gelu(self.conv6(x))
94
+ return x
95
+
96
+
97
+ class FeatureProjection(nn.Module):
98
+ def __init__(self):
99
+ super().__init__()
100
+ self.norm = nn.LayerNorm(512)
101
+ self.projection = nn.Linear(512, 768)
102
+ self.dropout = nn.Dropout(0.1)
103
+
104
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
105
+ x = self.norm(x)
106
+ x = self.projection(x)
107
+ x = self.dropout(x)
108
+ return x
109
+
110
+
111
+ class PositionalConvEmbedding(nn.Module):
112
+ def __init__(self):
113
+ super().__init__()
114
+ self.conv = nn.Conv1d(
115
+ 768,
116
+ 768,
117
+ kernel_size=128,
118
+ padding=128 // 2,
119
+ groups=16,
120
+ )
121
+ self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
122
+
123
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
124
+ x = self.conv(x.transpose(1, 2))
125
+ x = F.gelu(x[:, :, :-1])
126
+ return x.transpose(1, 2)
127
+
128
+
129
+ class TransformerEncoder(nn.Module):
130
+ def __init__(
131
+ self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
132
+ ) -> None:
133
+ super(TransformerEncoder, self).__init__()
134
+ self.layers = nn.ModuleList(
135
+ [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
136
+ )
137
+ self.num_layers = num_layers
138
+
139
+ def forward(
140
+ self,
141
+ src: torch.Tensor,
142
+ mask: torch.Tensor = None,
143
+ src_key_padding_mask: torch.Tensor = None,
144
+ output_layer: Optional[int] = None,
145
+ ) -> torch.Tensor:
146
+ output = src
147
+ for layer in self.layers[:output_layer]:
148
+ output = layer(
149
+ output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
150
+ )
151
+ return output
152
+
153
+
154
+ def _compute_mask(
155
+ shape: Tuple[int, int],
156
+ mask_prob: float,
157
+ mask_length: int,
158
+ device: torch.device,
159
+ min_masks: int = 0,
160
+ ) -> torch.Tensor:
161
+ batch_size, sequence_length = shape
162
+
163
+ if mask_length < 1:
164
+ raise ValueError("`mask_length` has to be bigger than 0.")
165
+
166
+ if mask_length > sequence_length:
167
+ raise ValueError(
168
+ f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
169
+ )
170
+
171
+ # compute number of masked spans in batch
172
+ num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
173
+ num_masked_spans = max(num_masked_spans, min_masks)
174
+
175
+ # make sure num masked indices <= sequence_length
176
+ if num_masked_spans * mask_length > sequence_length:
177
+ num_masked_spans = sequence_length // mask_length
178
+
179
+ # SpecAugment mask to fill
180
+ mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
181
+
182
+ # uniform distribution to sample from, make sure that offset samples are < sequence_length
183
+ uniform_dist = torch.ones(
184
+ (batch_size, sequence_length - (mask_length - 1)), device=device
185
+ )
186
+
187
+ # get random indices to mask
188
+ mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
189
+
190
+ # expand masked indices to masked spans
191
+ mask_indices = (
192
+ mask_indices.unsqueeze(dim=-1)
193
+ .expand((batch_size, num_masked_spans, mask_length))
194
+ .reshape(batch_size, num_masked_spans * mask_length)
195
+ )
196
+ offsets = (
197
+ torch.arange(mask_length, device=device)[None, None, :]
198
+ .expand((batch_size, num_masked_spans, mask_length))
199
+ .reshape(batch_size, num_masked_spans * mask_length)
200
+ )
201
+ mask_idxs = mask_indices + offsets
202
+
203
+ # scatter indices to mask
204
+ mask = mask.scatter(1, mask_idxs, True)
205
+
206
+ return mask
207
+
208
+
209
+ def hubert_soft(
210
+ path: str
211
+ ) -> HubertSoft:
212
+ r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
213
+ Args:
214
+ path (str): path of a pretrained model
215
+ """
216
+ hubert = HubertSoft()
217
+ checkpoint = torch.load(path)
218
+ consume_prefix_in_state_dict_if_present(checkpoint, "module.")
219
+ hubert.load_state_dict(checkpoint)
220
+ hubert.eval()
221
+ return hubert
mel_processing.py ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.utils.data
3
+ from librosa.filters import mel as librosa_mel_fn
4
+
5
+ MAX_WAV_VALUE = 32768.0
6
+
7
+
8
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
9
+ """
10
+ PARAMS
11
+ ------
12
+ C: compression factor
13
+ """
14
+ return torch.log(torch.clamp(x, min=clip_val) * C)
15
+
16
+
17
+ def dynamic_range_decompression_torch(x, C=1):
18
+ """
19
+ PARAMS
20
+ ------
21
+ C: compression factor used to compress
22
+ """
23
+ return torch.exp(x) / C
24
+
25
+
26
+ def spectral_normalize_torch(magnitudes):
27
+ output = dynamic_range_compression_torch(magnitudes)
28
+ return output
29
+
30
+
31
+ def spectral_de_normalize_torch(magnitudes):
32
+ output = dynamic_range_decompression_torch(magnitudes)
33
+ return output
34
+
35
+
36
+ mel_basis = {}
37
+ hann_window = {}
38
+
39
+
40
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
41
+ if torch.min(y) < -1.:
42
+ print('min value is ', torch.min(y))
43
+ if torch.max(y) > 1.:
44
+ print('max value is ', torch.max(y))
45
+
46
+ global hann_window
47
+ dtype_device = str(y.dtype) + '_' + str(y.device)
48
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
49
+ if wnsize_dtype_device not in hann_window:
50
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
51
+
52
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
53
+ y = y.squeeze(1)
54
+
55
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
56
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
57
+
58
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
59
+ return spec
60
+
61
+
62
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
63
+ global mel_basis
64
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
65
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
66
+ if fmax_dtype_device not in mel_basis:
67
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
68
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
69
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
70
+ spec = spectral_normalize_torch(spec)
71
+ return spec
72
+
73
+
74
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
75
+ if torch.min(y) < -1.:
76
+ print('min value is ', torch.min(y))
77
+ if torch.max(y) > 1.:
78
+ print('max value is ', torch.max(y))
79
+
80
+ global mel_basis, hann_window
81
+ dtype_device = str(y.dtype) + '_' + str(y.device)
82
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
83
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
84
+ if fmax_dtype_device not in mel_basis:
85
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
86
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
87
+ if wnsize_dtype_device not in hann_window:
88
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
89
+
90
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
91
+ y = y.squeeze(1)
92
+
93
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
94
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
95
+
96
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
97
+
98
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
99
+ spec = spectral_normalize_torch(spec)
100
+
101
+ return spec
models.py ADDED
@@ -0,0 +1,404 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ import commons
7
+ import modules
8
+ import attentions
9
+
10
+ from torch.nn import Conv1d, ConvTranspose1d
11
+ from torch.nn.utils import weight_norm
12
+ from commons import init_weights
13
+
14
+
15
+ class StochasticDurationPredictor(nn.Module):
16
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
17
+ super().__init__()
18
+ filter_channels = in_channels # it needs to be removed from future version.
19
+ self.in_channels = in_channels
20
+ self.filter_channels = filter_channels
21
+ self.kernel_size = kernel_size
22
+ self.p_dropout = p_dropout
23
+ self.n_flows = n_flows
24
+ self.gin_channels = gin_channels
25
+
26
+ self.log_flow = modules.Log()
27
+ self.flows = nn.ModuleList()
28
+ self.flows.append(modules.ElementwiseAffine(2))
29
+ for i in range(n_flows):
30
+ self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
31
+ self.flows.append(modules.Flip())
32
+
33
+ self.post_pre = nn.Conv1d(1, filter_channels, 1)
34
+ self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
35
+ self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
36
+ self.post_flows = nn.ModuleList()
37
+ self.post_flows.append(modules.ElementwiseAffine(2))
38
+ for i in range(4):
39
+ self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
40
+ self.post_flows.append(modules.Flip())
41
+
42
+ self.pre = nn.Conv1d(in_channels, filter_channels, 1)
43
+ self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
44
+ self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
45
+ if gin_channels != 0:
46
+ self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
47
+
48
+ def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
49
+ x = torch.detach(x)
50
+ x = self.pre(x)
51
+ if g is not None:
52
+ g = torch.detach(g)
53
+ x = x + self.cond(g)
54
+ x = self.convs(x, x_mask)
55
+ x = self.proj(x) * x_mask
56
+
57
+ if not reverse:
58
+ flows = self.flows
59
+ assert w is not None
60
+
61
+ logdet_tot_q = 0
62
+ h_w = self.post_pre(w)
63
+ h_w = self.post_convs(h_w, x_mask)
64
+ h_w = self.post_proj(h_w) * x_mask
65
+ e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
66
+ z_q = e_q
67
+ for flow in self.post_flows:
68
+ z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
69
+ logdet_tot_q += logdet_q
70
+ z_u, z1 = torch.split(z_q, [1, 1], 1)
71
+ u = torch.sigmoid(z_u) * x_mask
72
+ z0 = (w - u) * x_mask
73
+ logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])
74
+ logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q
75
+
76
+ logdet_tot = 0
77
+ z0, logdet = self.log_flow(z0, x_mask)
78
+ logdet_tot += logdet
79
+ z = torch.cat([z0, z1], 1)
80
+ for flow in flows:
81
+ z, logdet = flow(z, x_mask, g=x, reverse=reverse)
82
+ logdet_tot = logdet_tot + logdet
83
+ nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot
84
+ return nll + logq # [b]
85
+ else:
86
+ flows = list(reversed(self.flows))
87
+ flows = flows[:-2] + [flows[-1]] # remove a useless vflow
88
+ z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
89
+ for flow in flows:
90
+ z = flow(z, x_mask, g=x, reverse=reverse)
91
+ z0, z1 = torch.split(z, [1, 1], 1)
92
+ logw = z0
93
+ return logw
94
+
95
+
96
+ class DurationPredictor(nn.Module):
97
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
98
+ super().__init__()
99
+
100
+ self.in_channels = in_channels
101
+ self.filter_channels = filter_channels
102
+ self.kernel_size = kernel_size
103
+ self.p_dropout = p_dropout
104
+ self.gin_channels = gin_channels
105
+
106
+ self.drop = nn.Dropout(p_dropout)
107
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)
108
+ self.norm_1 = modules.LayerNorm(filter_channels)
109
+ self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
110
+ self.norm_2 = modules.LayerNorm(filter_channels)
111
+ self.proj = nn.Conv1d(filter_channels, 1, 1)
112
+
113
+ if gin_channels != 0:
114
+ self.cond = nn.Conv1d(gin_channels, in_channels, 1)
115
+
116
+ def forward(self, x, x_mask, g=None):
117
+ x = torch.detach(x)
118
+ if g is not None:
119
+ g = torch.detach(g)
120
+ x = x + self.cond(g)
121
+ x = self.conv_1(x * x_mask)
122
+ x = torch.relu(x)
123
+ x = self.norm_1(x)
124
+ x = self.drop(x)
125
+ x = self.conv_2(x * x_mask)
126
+ x = torch.relu(x)
127
+ x = self.norm_2(x)
128
+ x = self.drop(x)
129
+ x = self.proj(x * x_mask)
130
+ return x * x_mask
131
+
132
+
133
+ class TextEncoder(nn.Module):
134
+ def __init__(self,
135
+ n_vocab,
136
+ out_channels,
137
+ hidden_channels,
138
+ filter_channels,
139
+ n_heads,
140
+ n_layers,
141
+ kernel_size,
142
+ p_dropout,
143
+ emotion_embedding):
144
+ super().__init__()
145
+ self.n_vocab = n_vocab
146
+ self.out_channels = out_channels
147
+ self.hidden_channels = hidden_channels
148
+ self.filter_channels = filter_channels
149
+ self.n_heads = n_heads
150
+ self.n_layers = n_layers
151
+ self.kernel_size = kernel_size
152
+ self.p_dropout = p_dropout
153
+ self.emotion_embedding = emotion_embedding
154
+
155
+ if self.n_vocab!=0:
156
+ self.emb = nn.Embedding(n_vocab, hidden_channels)
157
+ if emotion_embedding:
158
+ self.emo_proj = nn.Linear(1024, hidden_channels)
159
+ nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
160
+
161
+ self.encoder = attentions.Encoder(
162
+ hidden_channels,
163
+ filter_channels,
164
+ n_heads,
165
+ n_layers,
166
+ kernel_size,
167
+ p_dropout)
168
+ self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
169
+
170
+ def forward(self, x, x_lengths, emotion_embedding=None):
171
+ if self.n_vocab!=0:
172
+ x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
173
+ if emotion_embedding is not None:
174
+ x = x + self.emo_proj(emotion_embedding.unsqueeze(1))
175
+ x = torch.transpose(x, 1, -1) # [b, h, t]
176
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
177
+
178
+ x = self.encoder(x * x_mask, x_mask)
179
+ stats = self.proj(x) * x_mask
180
+
181
+ m, logs = torch.split(stats, self.out_channels, dim=1)
182
+ return x, m, logs, x_mask
183
+
184
+
185
+ class ResidualCouplingBlock(nn.Module):
186
+ def __init__(self,
187
+ channels,
188
+ hidden_channels,
189
+ kernel_size,
190
+ dilation_rate,
191
+ n_layers,
192
+ n_flows=4,
193
+ gin_channels=0):
194
+ super().__init__()
195
+ self.channels = channels
196
+ self.hidden_channels = hidden_channels
197
+ self.kernel_size = kernel_size
198
+ self.dilation_rate = dilation_rate
199
+ self.n_layers = n_layers
200
+ self.n_flows = n_flows
201
+ self.gin_channels = gin_channels
202
+
203
+ self.flows = nn.ModuleList()
204
+ for i in range(n_flows):
205
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
206
+ self.flows.append(modules.Flip())
207
+
208
+ def forward(self, x, x_mask, g=None, reverse=False):
209
+ if not reverse:
210
+ for flow in self.flows:
211
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
212
+ else:
213
+ for flow in reversed(self.flows):
214
+ x = flow(x, x_mask, g=g, reverse=reverse)
215
+ return x
216
+
217
+
218
+ class PosteriorEncoder(nn.Module):
219
+ def __init__(self,
220
+ in_channels,
221
+ out_channels,
222
+ hidden_channels,
223
+ kernel_size,
224
+ dilation_rate,
225
+ n_layers,
226
+ gin_channels=0):
227
+ super().__init__()
228
+ self.in_channels = in_channels
229
+ self.out_channels = out_channels
230
+ self.hidden_channels = hidden_channels
231
+ self.kernel_size = kernel_size
232
+ self.dilation_rate = dilation_rate
233
+ self.n_layers = n_layers
234
+ self.gin_channels = gin_channels
235
+
236
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
237
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
238
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
239
+
240
+ def forward(self, x, x_lengths, g=None):
241
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
242
+ x = self.pre(x) * x_mask
243
+ x = self.enc(x, x_mask, g=g)
244
+ stats = self.proj(x) * x_mask
245
+ m, logs = torch.split(stats, self.out_channels, dim=1)
246
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
247
+ return z, m, logs, x_mask
248
+
249
+
250
+ class Generator(torch.nn.Module):
251
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
252
+ super(Generator, self).__init__()
253
+ self.num_kernels = len(resblock_kernel_sizes)
254
+ self.num_upsamples = len(upsample_rates)
255
+ self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
256
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
257
+
258
+ self.ups = nn.ModuleList()
259
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
260
+ self.ups.append(weight_norm(
261
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
262
+ k, u, padding=(k-u)//2)))
263
+
264
+ self.resblocks = nn.ModuleList()
265
+ for i in range(len(self.ups)):
266
+ ch = upsample_initial_channel//(2**(i+1))
267
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
268
+ self.resblocks.append(resblock(ch, k, d))
269
+
270
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
271
+ self.ups.apply(init_weights)
272
+
273
+ if gin_channels != 0:
274
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
275
+
276
+ def forward(self, x, g=None):
277
+ x = self.conv_pre(x)
278
+ if g is not None:
279
+ x = x + self.cond(g)
280
+
281
+ for i in range(self.num_upsamples):
282
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
283
+ x = self.ups[i](x)
284
+ xs = None
285
+ for j in range(self.num_kernels):
286
+ if xs is None:
287
+ xs = self.resblocks[i*self.num_kernels+j](x)
288
+ else:
289
+ xs += self.resblocks[i*self.num_kernels+j](x)
290
+ x = xs / self.num_kernels
291
+ x = F.leaky_relu(x)
292
+ x = self.conv_post(x)
293
+ x = torch.tanh(x)
294
+
295
+ return x
296
+
297
+
298
+ class SynthesizerTrn(nn.Module):
299
+ """
300
+ Synthesizer for Training
301
+ """
302
+
303
+ def __init__(self,
304
+ n_vocab,
305
+ spec_channels,
306
+ segment_size,
307
+ inter_channels,
308
+ hidden_channels,
309
+ filter_channels,
310
+ n_heads,
311
+ n_layers,
312
+ kernel_size,
313
+ p_dropout,
314
+ resblock,
315
+ resblock_kernel_sizes,
316
+ resblock_dilation_sizes,
317
+ upsample_rates,
318
+ upsample_initial_channel,
319
+ upsample_kernel_sizes,
320
+ n_speakers=0,
321
+ gin_channels=0,
322
+ use_sdp=True,
323
+ emotion_embedding=False,
324
+ **kwargs):
325
+
326
+ super().__init__()
327
+ self.n_vocab = n_vocab
328
+ self.spec_channels = spec_channels
329
+ self.inter_channels = inter_channels
330
+ self.hidden_channels = hidden_channels
331
+ self.filter_channels = filter_channels
332
+ self.n_heads = n_heads
333
+ self.n_layers = n_layers
334
+ self.kernel_size = kernel_size
335
+ self.p_dropout = p_dropout
336
+ self.resblock = resblock
337
+ self.resblock_kernel_sizes = resblock_kernel_sizes
338
+ self.resblock_dilation_sizes = resblock_dilation_sizes
339
+ self.upsample_rates = upsample_rates
340
+ self.upsample_initial_channel = upsample_initial_channel
341
+ self.upsample_kernel_sizes = upsample_kernel_sizes
342
+ self.segment_size = segment_size
343
+ self.n_speakers = n_speakers
344
+ self.gin_channels = gin_channels
345
+
346
+ self.use_sdp = use_sdp
347
+
348
+ self.enc_p = TextEncoder(n_vocab,
349
+ inter_channels,
350
+ hidden_channels,
351
+ filter_channels,
352
+ n_heads,
353
+ n_layers,
354
+ kernel_size,
355
+ p_dropout,
356
+ emotion_embedding)
357
+ self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
358
+ self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
359
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
360
+
361
+ if use_sdp:
362
+ self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
363
+ else:
364
+ self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
365
+
366
+ if n_speakers > 1:
367
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
368
+
369
+ def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None, emotion_embedding=None):
370
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, emotion_embedding)
371
+ if self.n_speakers > 0:
372
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
373
+ else:
374
+ g = None
375
+
376
+ if self.use_sdp:
377
+ logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
378
+ else:
379
+ logw = self.dp(x, x_mask, g=g)
380
+ w = torch.exp(logw) * x_mask * length_scale
381
+ w_ceil = torch.ceil(w)
382
+ y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
383
+ y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
384
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
385
+ attn = commons.generate_path(w_ceil, attn_mask)
386
+
387
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
388
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
389
+
390
+ z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
391
+ z = self.flow(z_p, y_mask, g=g, reverse=True)
392
+ o = self.dec((z * y_mask)[:,:,:max_len], g=g)
393
+ return o, attn, y_mask, (z, z_p, m_p, logs_p)
394
+
395
+ def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
396
+ assert self.n_speakers > 0, "n_speakers have to be larger than 0."
397
+ g_src = self.emb_g(sid_src).unsqueeze(-1)
398
+ g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
399
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
400
+ z_p = self.flow(z, y_mask, g=g_src)
401
+ z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
402
+ o_hat = self.dec(z_hat * y_mask, g=g_tgt)
403
+ return o_hat, y_mask, (z, z_p, z_hat)
404
+
modules.py ADDED
@@ -0,0 +1,387 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ from torch.nn import Conv1d
7
+ from torch.nn.utils import weight_norm, remove_weight_norm
8
+
9
+ import commons
10
+ from commons import init_weights, get_padding
11
+ from transforms import piecewise_rational_quadratic_transform
12
+
13
+
14
+ LRELU_SLOPE = 0.1
15
+
16
+
17
+ class LayerNorm(nn.Module):
18
+ def __init__(self, channels, eps=1e-5):
19
+ super().__init__()
20
+ self.channels = channels
21
+ self.eps = eps
22
+
23
+ self.gamma = nn.Parameter(torch.ones(channels))
24
+ self.beta = nn.Parameter(torch.zeros(channels))
25
+
26
+ def forward(self, x):
27
+ x = x.transpose(1, -1)
28
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
29
+ return x.transpose(1, -1)
30
+
31
+
32
+ class ConvReluNorm(nn.Module):
33
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
34
+ super().__init__()
35
+ self.in_channels = in_channels
36
+ self.hidden_channels = hidden_channels
37
+ self.out_channels = out_channels
38
+ self.kernel_size = kernel_size
39
+ self.n_layers = n_layers
40
+ self.p_dropout = p_dropout
41
+ assert n_layers > 1, "Number of layers should be larger than 0."
42
+
43
+ self.conv_layers = nn.ModuleList()
44
+ self.norm_layers = nn.ModuleList()
45
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
46
+ self.norm_layers.append(LayerNorm(hidden_channels))
47
+ self.relu_drop = nn.Sequential(
48
+ nn.ReLU(),
49
+ nn.Dropout(p_dropout))
50
+ for _ in range(n_layers-1):
51
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
52
+ self.norm_layers.append(LayerNorm(hidden_channels))
53
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
54
+ self.proj.weight.data.zero_()
55
+ self.proj.bias.data.zero_()
56
+
57
+ def forward(self, x, x_mask):
58
+ x_org = x
59
+ for i in range(self.n_layers):
60
+ x = self.conv_layers[i](x * x_mask)
61
+ x = self.norm_layers[i](x)
62
+ x = self.relu_drop(x)
63
+ x = x_org + self.proj(x)
64
+ return x * x_mask
65
+
66
+
67
+ class DDSConv(nn.Module):
68
+ """
69
+ Dilated and Depth-Separable Convolution
70
+ """
71
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
72
+ super().__init__()
73
+ self.channels = channels
74
+ self.kernel_size = kernel_size
75
+ self.n_layers = n_layers
76
+ self.p_dropout = p_dropout
77
+
78
+ self.drop = nn.Dropout(p_dropout)
79
+ self.convs_sep = nn.ModuleList()
80
+ self.convs_1x1 = nn.ModuleList()
81
+ self.norms_1 = nn.ModuleList()
82
+ self.norms_2 = nn.ModuleList()
83
+ for i in range(n_layers):
84
+ dilation = kernel_size ** i
85
+ padding = (kernel_size * dilation - dilation) // 2
86
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
87
+ groups=channels, dilation=dilation, padding=padding
88
+ ))
89
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
90
+ self.norms_1.append(LayerNorm(channels))
91
+ self.norms_2.append(LayerNorm(channels))
92
+
93
+ def forward(self, x, x_mask, g=None):
94
+ if g is not None:
95
+ x = x + g
96
+ for i in range(self.n_layers):
97
+ y = self.convs_sep[i](x * x_mask)
98
+ y = self.norms_1[i](y)
99
+ y = F.gelu(y)
100
+ y = self.convs_1x1[i](y)
101
+ y = self.norms_2[i](y)
102
+ y = F.gelu(y)
103
+ y = self.drop(y)
104
+ x = x + y
105
+ return x * x_mask
106
+
107
+
108
+ class WN(torch.nn.Module):
109
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
110
+ super(WN, self).__init__()
111
+ assert(kernel_size % 2 == 1)
112
+ self.hidden_channels =hidden_channels
113
+ self.kernel_size = kernel_size,
114
+ self.dilation_rate = dilation_rate
115
+ self.n_layers = n_layers
116
+ self.gin_channels = gin_channels
117
+ self.p_dropout = p_dropout
118
+
119
+ self.in_layers = torch.nn.ModuleList()
120
+ self.res_skip_layers = torch.nn.ModuleList()
121
+ self.drop = nn.Dropout(p_dropout)
122
+
123
+ if gin_channels != 0:
124
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
125
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
126
+
127
+ for i in range(n_layers):
128
+ dilation = dilation_rate ** i
129
+ padding = int((kernel_size * dilation - dilation) / 2)
130
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
131
+ dilation=dilation, padding=padding)
132
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
133
+ self.in_layers.append(in_layer)
134
+
135
+ # last one is not necessary
136
+ if i < n_layers - 1:
137
+ res_skip_channels = 2 * hidden_channels
138
+ else:
139
+ res_skip_channels = hidden_channels
140
+
141
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
142
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
143
+ self.res_skip_layers.append(res_skip_layer)
144
+
145
+ def forward(self, x, x_mask, g=None, **kwargs):
146
+ output = torch.zeros_like(x)
147
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
148
+
149
+ if g is not None:
150
+ g = self.cond_layer(g)
151
+
152
+ for i in range(self.n_layers):
153
+ x_in = self.in_layers[i](x)
154
+ if g is not None:
155
+ cond_offset = i * 2 * self.hidden_channels
156
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
157
+ else:
158
+ g_l = torch.zeros_like(x_in)
159
+
160
+ acts = commons.fused_add_tanh_sigmoid_multiply(
161
+ x_in,
162
+ g_l,
163
+ n_channels_tensor)
164
+ acts = self.drop(acts)
165
+
166
+ res_skip_acts = self.res_skip_layers[i](acts)
167
+ if i < self.n_layers - 1:
168
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
169
+ x = (x + res_acts) * x_mask
170
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
171
+ else:
172
+ output = output + res_skip_acts
173
+ return output * x_mask
174
+
175
+ def remove_weight_norm(self):
176
+ if self.gin_channels != 0:
177
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
178
+ for l in self.in_layers:
179
+ torch.nn.utils.remove_weight_norm(l)
180
+ for l in self.res_skip_layers:
181
+ torch.nn.utils.remove_weight_norm(l)
182
+
183
+
184
+ class ResBlock1(torch.nn.Module):
185
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
186
+ super(ResBlock1, self).__init__()
187
+ self.convs1 = nn.ModuleList([
188
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
189
+ padding=get_padding(kernel_size, dilation[0]))),
190
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
191
+ padding=get_padding(kernel_size, dilation[1]))),
192
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
193
+ padding=get_padding(kernel_size, dilation[2])))
194
+ ])
195
+ self.convs1.apply(init_weights)
196
+
197
+ self.convs2 = nn.ModuleList([
198
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
199
+ padding=get_padding(kernel_size, 1))),
200
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
201
+ padding=get_padding(kernel_size, 1))),
202
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
203
+ padding=get_padding(kernel_size, 1)))
204
+ ])
205
+ self.convs2.apply(init_weights)
206
+
207
+ def forward(self, x, x_mask=None):
208
+ for c1, c2 in zip(self.convs1, self.convs2):
209
+ xt = F.leaky_relu(x, LRELU_SLOPE)
210
+ if x_mask is not None:
211
+ xt = xt * x_mask
212
+ xt = c1(xt)
213
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
214
+ if x_mask is not None:
215
+ xt = xt * x_mask
216
+ xt = c2(xt)
217
+ x = xt + x
218
+ if x_mask is not None:
219
+ x = x * x_mask
220
+ return x
221
+
222
+ def remove_weight_norm(self):
223
+ for l in self.convs1:
224
+ remove_weight_norm(l)
225
+ for l in self.convs2:
226
+ remove_weight_norm(l)
227
+
228
+
229
+ class ResBlock2(torch.nn.Module):
230
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
231
+ super(ResBlock2, self).__init__()
232
+ self.convs = nn.ModuleList([
233
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
234
+ padding=get_padding(kernel_size, dilation[0]))),
235
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
236
+ padding=get_padding(kernel_size, dilation[1])))
237
+ ])
238
+ self.convs.apply(init_weights)
239
+
240
+ def forward(self, x, x_mask=None):
241
+ for c in self.convs:
242
+ xt = F.leaky_relu(x, LRELU_SLOPE)
243
+ if x_mask is not None:
244
+ xt = xt * x_mask
245
+ xt = c(xt)
246
+ x = xt + x
247
+ if x_mask is not None:
248
+ x = x * x_mask
249
+ return x
250
+
251
+ def remove_weight_norm(self):
252
+ for l in self.convs:
253
+ remove_weight_norm(l)
254
+
255
+
256
+ class Log(nn.Module):
257
+ def forward(self, x, x_mask, reverse=False, **kwargs):
258
+ if not reverse:
259
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
260
+ logdet = torch.sum(-y, [1, 2])
261
+ return y, logdet
262
+ else:
263
+ x = torch.exp(x) * x_mask
264
+ return x
265
+
266
+
267
+ class Flip(nn.Module):
268
+ def forward(self, x, *args, reverse=False, **kwargs):
269
+ x = torch.flip(x, [1])
270
+ if not reverse:
271
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
272
+ return x, logdet
273
+ else:
274
+ return x
275
+
276
+
277
+ class ElementwiseAffine(nn.Module):
278
+ def __init__(self, channels):
279
+ super().__init__()
280
+ self.channels = channels
281
+ self.m = nn.Parameter(torch.zeros(channels,1))
282
+ self.logs = nn.Parameter(torch.zeros(channels,1))
283
+
284
+ def forward(self, x, x_mask, reverse=False, **kwargs):
285
+ if not reverse:
286
+ y = self.m + torch.exp(self.logs) * x
287
+ y = y * x_mask
288
+ logdet = torch.sum(self.logs * x_mask, [1,2])
289
+ return y, logdet
290
+ else:
291
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
292
+ return x
293
+
294
+
295
+ class ResidualCouplingLayer(nn.Module):
296
+ def __init__(self,
297
+ channels,
298
+ hidden_channels,
299
+ kernel_size,
300
+ dilation_rate,
301
+ n_layers,
302
+ p_dropout=0,
303
+ gin_channels=0,
304
+ mean_only=False):
305
+ assert channels % 2 == 0, "channels should be divisible by 2"
306
+ super().__init__()
307
+ self.channels = channels
308
+ self.hidden_channels = hidden_channels
309
+ self.kernel_size = kernel_size
310
+ self.dilation_rate = dilation_rate
311
+ self.n_layers = n_layers
312
+ self.half_channels = channels // 2
313
+ self.mean_only = mean_only
314
+
315
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
316
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
317
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
318
+ self.post.weight.data.zero_()
319
+ self.post.bias.data.zero_()
320
+
321
+ def forward(self, x, x_mask, g=None, reverse=False):
322
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
323
+ h = self.pre(x0) * x_mask
324
+ h = self.enc(h, x_mask, g=g)
325
+ stats = self.post(h) * x_mask
326
+ if not self.mean_only:
327
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
328
+ else:
329
+ m = stats
330
+ logs = torch.zeros_like(m)
331
+
332
+ if not reverse:
333
+ x1 = m + x1 * torch.exp(logs) * x_mask
334
+ x = torch.cat([x0, x1], 1)
335
+ logdet = torch.sum(logs, [1,2])
336
+ return x, logdet
337
+ else:
338
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
339
+ x = torch.cat([x0, x1], 1)
340
+ return x
341
+
342
+
343
+ class ConvFlow(nn.Module):
344
+ def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
345
+ super().__init__()
346
+ self.in_channels = in_channels
347
+ self.filter_channels = filter_channels
348
+ self.kernel_size = kernel_size
349
+ self.n_layers = n_layers
350
+ self.num_bins = num_bins
351
+ self.tail_bound = tail_bound
352
+ self.half_channels = in_channels // 2
353
+
354
+ self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
355
+ self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
356
+ self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
357
+ self.proj.weight.data.zero_()
358
+ self.proj.bias.data.zero_()
359
+
360
+ def forward(self, x, x_mask, g=None, reverse=False):
361
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
362
+ h = self.pre(x0)
363
+ h = self.convs(h, x_mask, g=g)
364
+ h = self.proj(h) * x_mask
365
+
366
+ b, c, t = x0.shape
367
+ h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
368
+
369
+ unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
370
+ unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
371
+ unnormalized_derivatives = h[..., 2 * self.num_bins:]
372
+
373
+ x1, logabsdet = piecewise_rational_quadratic_transform(x1,
374
+ unnormalized_widths,
375
+ unnormalized_heights,
376
+ unnormalized_derivatives,
377
+ inverse=reverse,
378
+ tails='linear',
379
+ tail_bound=self.tail_bound
380
+ )
381
+
382
+ x = torch.cat([x0, x1], 1) * x_mask
383
+ logdet = torch.sum(logabsdet * x_mask, [1,2])
384
+ if not reverse:
385
+ return x, logdet
386
+ else:
387
+ return x
requirements.txt ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ numba
2
+ librosa
3
+ numpy==1.23.3
4
+ scipy
5
+ torch
6
+ unidecode
7
+ openjtalk>=0.3.0.dev2
8
+ jamo
9
+ pypinyin
10
+ jieba
11
+ protobuf
12
+ cn2an
13
+ inflect
14
+ eng_to_ipa
15
+ ko_pron
16
+ indic_transliteration
17
+ num_thai
18
+ opencc
19
+ audonnx
20
+ flask
21
+ av
22
+ pilk
23
+ flask_apscheduler
text/LICENSE ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright (c) 2017 Keith Ito
2
+
3
+ Permission is hereby granted, free of charge, to any person obtaining a copy
4
+ of this software and associated documentation files (the "Software"), to deal
5
+ in the Software without restriction, including without limitation the rights
6
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7
+ copies of the Software, and to permit persons to whom the Software is
8
+ furnished to do so, subject to the following conditions:
9
+
10
+ The above copyright notice and this permission notice shall be included in
11
+ all copies or substantial portions of the Software.
12
+
13
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
19
+ THE SOFTWARE.
text/__init__.py ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+ from text import cleaners
3
+
4
+
5
+ def text_to_sequence(text, symbols, cleaner_names):
6
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
7
+ Args:
8
+ text: string to convert to a sequence
9
+ cleaner_names: names of the cleaner functions to run the text through
10
+ Returns:
11
+ List of integers corresponding to the symbols in the text
12
+ '''
13
+ _symbol_to_id = {s: i for i, s in enumerate(symbols)}
14
+
15
+ sequence = []
16
+
17
+ clean_text = _clean_text(text, cleaner_names)
18
+ for symbol in clean_text:
19
+ if symbol not in _symbol_to_id.keys():
20
+ continue
21
+ symbol_id = _symbol_to_id[symbol]
22
+ sequence += [symbol_id]
23
+ return sequence
24
+
25
+
26
+ def _clean_text(text, cleaner_names):
27
+ for name in cleaner_names:
28
+ cleaner = getattr(cleaners, name)
29
+ if not cleaner:
30
+ raise Exception('Unknown cleaner: %s' % name)
31
+ text = cleaner(text)
32
+ return text
text/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (1.21 kB). View file
 
text/__pycache__/cleaners.cpython-310.pyc ADDED
Binary file (6.85 kB). View file
 
text/__pycache__/japanese.cpython-310.pyc ADDED
Binary file (4.13 kB). View file
 
text/__pycache__/mandarin.cpython-310.pyc ADDED
Binary file (6.17 kB). View file
 
text/cantonese.py ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ import cn2an
3
+ import opencc
4
+
5
+
6
+ converter = opencc.OpenCC('jyutjyu')
7
+
8
+ # List of (Latin alphabet, ipa) pairs:
9
+ _latin_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
10
+ ('A', 'ei˥'),
11
+ ('B', 'biː˥'),
12
+ ('C', 'siː˥'),
13
+ ('D', 'tiː˥'),
14
+ ('E', 'iː˥'),
15
+ ('F', 'e˥fuː˨˩'),
16
+ ('G', 'tsiː˥'),
17
+ ('H', 'ɪk̚˥tsʰyː˨˩'),
18
+ ('I', 'ɐi˥'),
19
+ ('J', 'tsei˥'),
20
+ ('K', 'kʰei˥'),
21
+ ('L', 'e˥llou˨˩'),
22
+ ('M', 'ɛːm˥'),
23
+ ('N', 'ɛːn˥'),
24
+ ('O', 'ou˥'),
25
+ ('P', 'pʰiː˥'),
26
+ ('Q', 'kʰiːu˥'),
27
+ ('R', 'aː˥lou˨˩'),
28
+ ('S', 'ɛː˥siː˨˩'),
29
+ ('T', 'tʰiː˥'),
30
+ ('U', 'juː˥'),
31
+ ('V', 'wiː˥'),
32
+ ('W', 'tʊk̚˥piː˥juː˥'),
33
+ ('X', 'ɪk̚˥siː˨˩'),
34
+ ('Y', 'waːi˥'),
35
+ ('Z', 'iː˨sɛːt̚˥')
36
+ ]]
37
+
38
+
39
+ def number_to_cantonese(text):
40
+ return re.sub(r'\d+(?:\.?\d+)?', lambda x: cn2an.an2cn(x.group()), text)
41
+
42
+
43
+ def latin_to_ipa(text):
44
+ for regex, replacement in _latin_to_ipa:
45
+ text = re.sub(regex, replacement, text)
46
+ return text
47
+
48
+
49
+ def cantonese_to_ipa(text):
50
+ text = number_to_cantonese(text.upper())
51
+ text = converter.convert(text).replace('-','').replace('$',' ')
52
+ text = re.sub(r'[A-Z]', lambda x: latin_to_ipa(x.group())+' ', text)
53
+ text = re.sub(r'[、;:]', ',', text)
54
+ text = re.sub(r'\s*,\s*', ', ', text)
55
+ text = re.sub(r'\s*。\s*', '. ', text)
56
+ text = re.sub(r'\s*?\s*', '? ', text)
57
+ text = re.sub(r'\s*!\s*', '! ', text)
58
+ text = re.sub(r'\s*$', '', text)
59
+ return text
text/cleaners.py ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+
3
+
4
+ def japanese_cleaners(text):
5
+ from text.japanese import japanese_to_romaji_with_accent
6
+ text = japanese_to_romaji_with_accent(text)
7
+ text = re.sub(r'([A-Za-z])$', r'\1.', text)
8
+ return text
9
+
10
+
11
+ def japanese_cleaners2(text):
12
+ return japanese_cleaners(text).replace('ts', 'ʦ').replace('...', '…')
13
+
14
+
15
+ def korean_cleaners(text):
16
+ '''Pipeline for Korean text'''
17
+ from text.korean import latin_to_hangul, number_to_hangul, divide_hangul
18
+ text = latin_to_hangul(text)
19
+ text = number_to_hangul(text)
20
+ text = divide_hangul(text)
21
+ text = re.sub(r'([\u3131-\u3163])$', r'\1.', text)
22
+ return text
23
+
24
+
25
+ def chinese_cleaners(text):
26
+ '''Pipeline for Chinese text'''
27
+ from text.mandarin import number_to_chinese, chinese_to_bopomofo, latin_to_bopomofo
28
+ text = number_to_chinese(text)
29
+ text = chinese_to_bopomofo(text)
30
+ text = latin_to_bopomofo(text)
31
+ text = re.sub(r'([ˉˊˇˋ˙])$', r'\1。', text)
32
+ return text
33
+
34
+
35
+ def zh_ja_mixture_cleaners(text):
36
+ from text.mandarin import chinese_to_romaji
37
+ from text.japanese import japanese_to_romaji_with_accent
38
+ text = re.sub(r'\[ZH\](.*?)\[ZH\]',
39
+ lambda x: chinese_to_romaji(x.group(1))+' ', text)
40
+ text = re.sub(r'\[JA\](.*?)\[JA\]', lambda x: japanese_to_romaji_with_accent(
41
+ x.group(1)).replace('ts', 'ʦ').replace('u', 'ɯ').replace('...', '…')+' ', text)
42
+ text = re.sub(r'\s+$', '', text)
43
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
44
+ return text
45
+
46
+
47
+ def sanskrit_cleaners(text):
48
+ text = text.replace('॥', '।').replace('ॐ', 'ओम्')
49
+ text = re.sub(r'([^।])$', r'\1।', text)
50
+ return text
51
+
52
+
53
+ def cjks_cleaners(text):
54
+ from text.mandarin import chinese_to_lazy_ipa
55
+ from text.japanese import japanese_to_ipa
56
+ from text.korean import korean_to_lazy_ipa
57
+ from text.sanskrit import devanagari_to_ipa
58
+ from text.english import english_to_lazy_ipa
59
+ text = re.sub(r'\[ZH\](.*?)\[ZH\]',
60
+ lambda x: chinese_to_lazy_ipa(x.group(1))+' ', text)
61
+ text = re.sub(r'\[JA\](.*?)\[JA\]',
62
+ lambda x: japanese_to_ipa(x.group(1))+' ', text)
63
+ text = re.sub(r'\[KO\](.*?)\[KO\]',
64
+ lambda x: korean_to_lazy_ipa(x.group(1))+' ', text)
65
+ text = re.sub(r'\[SA\](.*?)\[SA\]',
66
+ lambda x: devanagari_to_ipa(x.group(1))+' ', text)
67
+ text = re.sub(r'\[EN\](.*?)\[EN\]',
68
+ lambda x: english_to_lazy_ipa(x.group(1))+' ', text)
69
+ text = re.sub(r'\s+$', '', text)
70
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
71
+ return text
72
+
73
+
74
+ def cjke_cleaners(text):
75
+ from text.mandarin import chinese_to_lazy_ipa
76
+ from text.japanese import japanese_to_ipa
77
+ from text.korean import korean_to_ipa
78
+ from text.english import english_to_ipa2
79
+ text = re.sub(r'\[ZH\](.*?)\[ZH\]', lambda x: chinese_to_lazy_ipa(x.group(1)).replace(
80
+ 'ʧ', 'tʃ').replace('ʦ', 'ts').replace('ɥan', 'ɥæn')+' ', text)
81
+ text = re.sub(r'\[JA\](.*?)\[JA\]', lambda x: japanese_to_ipa(x.group(1)).replace('ʧ', 'tʃ').replace(
82
+ 'ʦ', 'ts').replace('ɥan', 'ɥæn').replace('ʥ', 'dz')+' ', text)
83
+ text = re.sub(r'\[KO\](.*?)\[KO\]',
84
+ lambda x: korean_to_ipa(x.group(1))+' ', text)
85
+ text = re.sub(r'\[EN\](.*?)\[EN\]', lambda x: english_to_ipa2(x.group(1)).replace('ɑ', 'a').replace(
86
+ 'ɔ', 'o').replace('ɛ', 'e').replace('ɪ', 'i').replace('ʊ', 'u')+' ', text)
87
+ text = re.sub(r'\s+$', '', text)
88
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
89
+ return text
90
+
91
+
92
+ def cjke_cleaners2(text):
93
+ from text.mandarin import chinese_to_ipa
94
+ from text.japanese import japanese_to_ipa2
95
+ from text.korean import korean_to_ipa
96
+ from text.english import english_to_ipa2
97
+ text = re.sub(r'\[ZH\](.*?)\[ZH\]',
98
+ lambda x: chinese_to_ipa(x.group(1))+' ', text)
99
+ text = re.sub(r'\[JA\](.*?)\[JA\]',
100
+ lambda x: japanese_to_ipa2(x.group(1))+' ', text)
101
+ text = re.sub(r'\[KO\](.*?)\[KO\]',
102
+ lambda x: korean_to_ipa(x.group(1))+' ', text)
103
+ text = re.sub(r'\[EN\](.*?)\[EN\]',
104
+ lambda x: english_to_ipa2(x.group(1))+' ', text)
105
+ text = re.sub(r'\s+$', '', text)
106
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
107
+ return text
108
+
109
+
110
+ def thai_cleaners(text):
111
+ from text.thai import num_to_thai, latin_to_thai
112
+ text = num_to_thai(text)
113
+ text = latin_to_thai(text)
114
+ return text
115
+
116
+
117
+ def shanghainese_cleaners(text):
118
+ from text.shanghainese import shanghainese_to_ipa
119
+ text = shanghainese_to_ipa(text)
120
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
121
+ return text
122
+
123
+
124
+ def chinese_dialect_cleaners(text):
125
+ from text.mandarin import chinese_to_ipa2
126
+ from text.japanese import japanese_to_ipa3
127
+ from text.shanghainese import shanghainese_to_ipa
128
+ from text.cantonese import cantonese_to_ipa
129
+ from text.english import english_to_lazy_ipa2
130
+ from text.ngu_dialect import ngu_dialect_to_ipa
131
+ text = re.sub(r'\[ZH\](.*?)\[ZH\]',
132
+ lambda x: chinese_to_ipa2(x.group(1))+' ', text)
133
+ text = re.sub(r'\[JA\](.*?)\[JA\]',
134
+ lambda x: japanese_to_ipa3(x.group(1)).replace('Q', 'ʔ')+' ', text)
135
+ text = re.sub(r'\[SH\](.*?)\[SH\]', lambda x: shanghainese_to_ipa(x.group(1)).replace('1', '˥˧').replace('5',
136
+ '˧˧˦').replace('6', '˩˩˧').replace('7', '˥').replace('8', '˩˨').replace('ᴀ', 'ɐ').replace('ᴇ', 'e')+' ', text)
137
+ text = re.sub(r'\[GD\](.*?)\[GD\]',
138
+ lambda x: cantonese_to_ipa(x.group(1))+' ', text)
139
+ text = re.sub(r'\[EN\](.*?)\[EN\]',
140
+ lambda x: english_to_lazy_ipa2(x.group(1))+' ', text)
141
+ text = re.sub(r'\[([A-Z]{2})\](.*?)\[\1\]', lambda x: ngu_dialect_to_ipa(x.group(2), x.group(
142
+ 1)).replace('ʣ', 'dz').replace('ʥ', 'dʑ').replace('ʦ', 'ts').replace('ʨ', 'tɕ')+' ', text)
143
+ text = re.sub(r'\s+$', '', text)
144
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
145
+ return text
text/english.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+
3
+ '''
4
+ Cleaners are transformations that run over the input text at both training and eval time.
5
+
6
+ Cleaners can be selected by passing a comma-delimited list of cleaner names as the "cleaners"
7
+ hyperparameter. Some cleaners are English-specific. You'll typically want to use:
8
+ 1. "english_cleaners" for English text
9
+ 2. "transliteration_cleaners" for non-English text that can be transliterated to ASCII using
10
+ the Unidecode library (https://pypi.python.org/pypi/Unidecode)
11
+ 3. "basic_cleaners" if you do not want to transliterate (in this case, you should also update
12
+ the symbols in symbols.py to match your data).
13
+ '''
14
+
15
+
16
+ # Regular expression matching whitespace:
17
+
18
+
19
+ import re
20
+ import inflect
21
+ from unidecode import unidecode
22
+ import eng_to_ipa as ipa
23
+ _inflect = inflect.engine()
24
+ _comma_number_re = re.compile(r'([0-9][0-9\,]+[0-9])')
25
+ _decimal_number_re = re.compile(r'([0-9]+\.[0-9]+)')
26
+ _pounds_re = re.compile(r'£([0-9\,]*[0-9]+)')
27
+ _dollars_re = re.compile(r'\$([0-9\.\,]*[0-9]+)')
28
+ _ordinal_re = re.compile(r'[0-9]+(st|nd|rd|th)')
29
+ _number_re = re.compile(r'[0-9]+')
30
+
31
+ # List of (regular expression, replacement) pairs for abbreviations:
32
+ _abbreviations = [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1]) for x in [
33
+ ('mrs', 'misess'),
34
+ ('mr', 'mister'),
35
+ ('dr', 'doctor'),
36
+ ('st', 'saint'),
37
+ ('co', 'company'),
38
+ ('jr', 'junior'),
39
+ ('maj', 'major'),
40
+ ('gen', 'general'),
41
+ ('drs', 'doctors'),
42
+ ('rev', 'reverend'),
43
+ ('lt', 'lieutenant'),
44
+ ('hon', 'honorable'),
45
+ ('sgt', 'sergeant'),
46
+ ('capt', 'captain'),
47
+ ('esq', 'esquire'),
48
+ ('ltd', 'limited'),
49
+ ('col', 'colonel'),
50
+ ('ft', 'fort'),
51
+ ]]
52
+
53
+
54
+ # List of (ipa, lazy ipa) pairs:
55
+ _lazy_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
56
+ ('r', 'ɹ'),
57
+ ('æ', 'e'),
58
+ ('ɑ', 'a'),
59
+ ('ɔ', 'o'),
60
+ ('ð', 'z'),
61
+ ('θ', 's'),
62
+ ('ɛ', 'e'),
63
+ ('ɪ', 'i'),
64
+ ('ʊ', 'u'),
65
+ ('ʒ', 'ʥ'),
66
+ ('ʤ', 'ʥ'),
67
+ ('ˈ', '↓'),
68
+ ]]
69
+
70
+ # List of (ipa, lazy ipa2) pairs:
71
+ _lazy_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
72
+ ('r', 'ɹ'),
73
+ ('ð', 'z'),
74
+ ('θ', 's'),
75
+ ('ʒ', 'ʑ'),
76
+ ('ʤ', 'dʑ'),
77
+ ('ˈ', '↓'),
78
+ ]]
79
+
80
+ # List of (ipa, ipa2) pairs
81
+ _ipa_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
82
+ ('r', 'ɹ'),
83
+ ('ʤ', 'dʒ'),
84
+ ('ʧ', 'tʃ')
85
+ ]]
86
+
87
+
88
+ def expand_abbreviations(text):
89
+ for regex, replacement in _abbreviations:
90
+ text = re.sub(regex, replacement, text)
91
+ return text
92
+
93
+
94
+ def collapse_whitespace(text):
95
+ return re.sub(r'\s+', ' ', text)
96
+
97
+
98
+ def _remove_commas(m):
99
+ return m.group(1).replace(',', '')
100
+
101
+
102
+ def _expand_decimal_point(m):
103
+ return m.group(1).replace('.', ' point ')
104
+
105
+
106
+ def _expand_dollars(m):
107
+ match = m.group(1)
108
+ parts = match.split('.')
109
+ if len(parts) > 2:
110
+ return match + ' dollars' # Unexpected format
111
+ dollars = int(parts[0]) if parts[0] else 0
112
+ cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
113
+ if dollars and cents:
114
+ dollar_unit = 'dollar' if dollars == 1 else 'dollars'
115
+ cent_unit = 'cent' if cents == 1 else 'cents'
116
+ return '%s %s, %s %s' % (dollars, dollar_unit, cents, cent_unit)
117
+ elif dollars:
118
+ dollar_unit = 'dollar' if dollars == 1 else 'dollars'
119
+ return '%s %s' % (dollars, dollar_unit)
120
+ elif cents:
121
+ cent_unit = 'cent' if cents == 1 else 'cents'
122
+ return '%s %s' % (cents, cent_unit)
123
+ else:
124
+ return 'zero dollars'
125
+
126
+
127
+ def _expand_ordinal(m):
128
+ return _inflect.number_to_words(m.group(0))
129
+
130
+
131
+ def _expand_number(m):
132
+ num = int(m.group(0))
133
+ if num > 1000 and num < 3000:
134
+ if num == 2000:
135
+ return 'two thousand'
136
+ elif num > 2000 and num < 2010:
137
+ return 'two thousand ' + _inflect.number_to_words(num % 100)
138
+ elif num % 100 == 0:
139
+ return _inflect.number_to_words(num // 100) + ' hundred'
140
+ else:
141
+ return _inflect.number_to_words(num, andword='', zero='oh', group=2).replace(', ', ' ')
142
+ else:
143
+ return _inflect.number_to_words(num, andword='')
144
+
145
+
146
+ def normalize_numbers(text):
147
+ text = re.sub(_comma_number_re, _remove_commas, text)
148
+ text = re.sub(_pounds_re, r'\1 pounds', text)
149
+ text = re.sub(_dollars_re, _expand_dollars, text)
150
+ text = re.sub(_decimal_number_re, _expand_decimal_point, text)
151
+ text = re.sub(_ordinal_re, _expand_ordinal, text)
152
+ text = re.sub(_number_re, _expand_number, text)
153
+ return text
154
+
155
+
156
+ def mark_dark_l(text):
157
+ return re.sub(r'l([^aeiouæɑɔəɛɪʊ ]*(?: |$))', lambda x: 'ɫ'+x.group(1), text)
158
+
159
+
160
+ def english_to_ipa(text):
161
+ text = unidecode(text).lower()
162
+ text = expand_abbreviations(text)
163
+ text = normalize_numbers(text)
164
+ phonemes = ipa.convert(text)
165
+ phonemes = collapse_whitespace(phonemes)
166
+ return phonemes
167
+
168
+
169
+ def english_to_lazy_ipa(text):
170
+ text = english_to_ipa(text)
171
+ for regex, replacement in _lazy_ipa:
172
+ text = re.sub(regex, replacement, text)
173
+ return text
174
+
175
+
176
+ def english_to_ipa2(text):
177
+ text = english_to_ipa(text)
178
+ text = mark_dark_l(text)
179
+ for regex, replacement in _ipa_to_ipa2:
180
+ text = re.sub(regex, replacement, text)
181
+ return text.replace('...', '…')
182
+
183
+
184
+ def english_to_lazy_ipa2(text):
185
+ text = english_to_ipa(text)
186
+ for regex, replacement in _lazy_ipa2:
187
+ text = re.sub(regex, replacement, text)
188
+ return text
text/japanese.py ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from unidecode import unidecode
3
+ import pyopenjtalk
4
+
5
+
6
+ # Regular expression matching Japanese without punctuation marks:
7
+ _japanese_characters = re.compile(
8
+ r'[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
9
+
10
+ # Regular expression matching non-Japanese characters or punctuation marks:
11
+ _japanese_marks = re.compile(
12
+ r'[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
13
+
14
+ # List of (symbol, Japanese) pairs for marks:
15
+ _symbols_to_japanese = [(re.compile('%s' % x[0]), x[1]) for x in [
16
+ ('%', 'パーセント')
17
+ ]]
18
+
19
+ # List of (romaji, ipa) pairs for marks:
20
+ _romaji_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
21
+ ('ts', 'ʦ'),
22
+ ('u', 'ɯ'),
23
+ ('j', 'ʥ'),
24
+ ('y', 'j'),
25
+ ('ni', 'n^i'),
26
+ ('nj', 'n^'),
27
+ ('hi', 'çi'),
28
+ ('hj', 'ç'),
29
+ ('f', 'ɸ'),
30
+ ('I', 'i*'),
31
+ ('U', 'ɯ*'),
32
+ ('r', 'ɾ')
33
+ ]]
34
+
35
+ # List of (romaji, ipa2) pairs for marks:
36
+ _romaji_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
37
+ ('u', 'ɯ'),
38
+ ('ʧ', 'tʃ'),
39
+ ('j', 'dʑ'),
40
+ ('y', 'j'),
41
+ ('ni', 'n^i'),
42
+ ('nj', 'n^'),
43
+ ('hi', 'çi'),
44
+ ('hj', 'ç'),
45
+ ('f', 'ɸ'),
46
+ ('I', 'i*'),
47
+ ('U', 'ɯ*'),
48
+ ('r', 'ɾ')
49
+ ]]
50
+
51
+ # List of (consonant, sokuon) pairs:
52
+ _real_sokuon = [(re.compile('%s' % x[0]), x[1]) for x in [
53
+ (r'Q([↑↓]*[kg])', r'k#\1'),
54
+ (r'Q([↑↓]*[tdjʧ])', r't#\1'),
55
+ (r'Q([↑↓]*[sʃ])', r's\1'),
56
+ (r'Q([↑↓]*[pb])', r'p#\1')
57
+ ]]
58
+
59
+ # List of (consonant, hatsuon) pairs:
60
+ _real_hatsuon = [(re.compile('%s' % x[0]), x[1]) for x in [
61
+ (r'N([↑↓]*[pbm])', r'm\1'),
62
+ (r'N([↑↓]*[ʧʥj])', r'n^\1'),
63
+ (r'N([↑↓]*[tdn])', r'n\1'),
64
+ (r'N([↑↓]*[kg])', r'ŋ\1')
65
+ ]]
66
+
67
+
68
+ def symbols_to_japanese(text):
69
+ for regex, replacement in _symbols_to_japanese:
70
+ text = re.sub(regex, replacement, text)
71
+ return text
72
+
73
+
74
+ def japanese_to_romaji_with_accent(text):
75
+ '''Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html'''
76
+ text = symbols_to_japanese(text)
77
+ sentences = re.split(_japanese_marks, text)
78
+ marks = re.findall(_japanese_marks, text)
79
+ text = ''
80
+ for i, sentence in enumerate(sentences):
81
+ if re.match(_japanese_characters, sentence):
82
+ if text != '':
83
+ text += ' '
84
+ labels = pyopenjtalk.extract_fullcontext(sentence)
85
+ for n, label in enumerate(labels):
86
+ phoneme = re.search(r'\-([^\+]*)\+', label).group(1)
87
+ if phoneme not in ['sil', 'pau']:
88
+ text += phoneme.replace('ch', 'ʧ').replace('sh',
89
+ 'ʃ').replace('cl', 'Q')
90
+ else:
91
+ continue
92
+ # n_moras = int(re.search(r'/F:(\d+)_', label).group(1))
93
+ a1 = int(re.search(r"/A:(\-?[0-9]+)\+", label).group(1))
94
+ a2 = int(re.search(r"\+(\d+)\+", label).group(1))
95
+ a3 = int(re.search(r"\+(\d+)/", label).group(1))
96
+ if re.search(r'\-([^\+]*)\+', labels[n + 1]).group(1) in ['sil', 'pau']:
97
+ a2_next = -1
98
+ else:
99
+ a2_next = int(
100
+ re.search(r"\+(\d+)\+", labels[n + 1]).group(1))
101
+ # Accent phrase boundary
102
+ if a3 == 1 and a2_next == 1:
103
+ text += ' '
104
+ # Falling
105
+ elif a1 == 0 and a2_next == a2 + 1:
106
+ text += '↓'
107
+ # Rising
108
+ elif a2 == 1 and a2_next == 2:
109
+ text += '↑'
110
+ if i < len(marks):
111
+ text += unidecode(marks[i]).replace(' ', '')
112
+ return text
113
+
114
+
115
+ def get_real_sokuon(text):
116
+ for regex, replacement in _real_sokuon:
117
+ text = re.sub(regex, replacement, text)
118
+ return text
119
+
120
+
121
+ def get_real_hatsuon(text):
122
+ for regex, replacement in _real_hatsuon:
123
+ text = re.sub(regex, replacement, text)
124
+ return text
125
+
126
+
127
+ def japanese_to_ipa(text):
128
+ text = japanese_to_romaji_with_accent(text).replace('...', '…')
129
+ text = re.sub(
130
+ r'([aiueo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
131
+ text = get_real_sokuon(text)
132
+ text = get_real_hatsuon(text)
133
+ for regex, replacement in _romaji_to_ipa:
134
+ text = re.sub(regex, replacement, text)
135
+ return text
136
+
137
+
138
+ def japanese_to_ipa2(text):
139
+ text = japanese_to_romaji_with_accent(text).replace('...', '…')
140
+ text = get_real_sokuon(text)
141
+ text = get_real_hatsuon(text)
142
+ for regex, replacement in _romaji_to_ipa2:
143
+ text = re.sub(regex, replacement, text)
144
+ return text
145
+
146
+
147
+ def japanese_to_ipa3(text):
148
+ text = japanese_to_ipa2(text).replace('n^', 'ȵ').replace(
149
+ 'ʃ', 'ɕ').replace('*', '\u0325').replace('#', '\u031a')
150
+ text = re.sub(
151
+ r'([aiɯeo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
152
+ text = re.sub(r'((?:^|\s)(?:ts|tɕ|[kpt]))', r'\1ʰ', text)
153
+ return text
text/korean.py ADDED
@@ -0,0 +1,210 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from jamo import h2j, j2hcj
3
+ import ko_pron
4
+
5
+
6
+ # This is a list of Korean classifiers preceded by pure Korean numerals.
7
+ _korean_classifiers = '군데 권 개 그루 닢 대 두 마리 모 모금 뭇 발 발짝 방 번 벌 보루 살 수 술 시 쌈 움큼 정 짝 채 척 첩 축 켤레 톨 통'
8
+
9
+ # List of (hangul, hangul divided) pairs:
10
+ _hangul_divided = [(re.compile('%s' % x[0]), x[1]) for x in [
11
+ ('ㄳ', 'ㄱㅅ'),
12
+ ('ㄵ', 'ㄴㅈ'),
13
+ ('ㄶ', 'ㄴㅎ'),
14
+ ('ㄺ', 'ㄹㄱ'),
15
+ ('ㄻ', 'ㄹㅁ'),
16
+ ('ㄼ', 'ㄹㅂ'),
17
+ ('ㄽ', 'ㄹㅅ'),
18
+ ('ㄾ', 'ㄹㅌ'),
19
+ ('ㄿ', 'ㄹㅍ'),
20
+ ('ㅀ', 'ㄹㅎ'),
21
+ ('ㅄ', 'ㅂㅅ'),
22
+ ('ㅘ', 'ㅗㅏ'),
23
+ ('ㅙ', 'ㅗㅐ'),
24
+ ('ㅚ', 'ㅗㅣ'),
25
+ ('ㅝ', 'ㅜㅓ'),
26
+ ('ㅞ', 'ㅜㅔ'),
27
+ ('ㅟ', 'ㅜㅣ'),
28
+ ('ㅢ', 'ㅡㅣ'),
29
+ ('ㅑ', 'ㅣㅏ'),
30
+ ('ㅒ', 'ㅣㅐ'),
31
+ ('ㅕ', 'ㅣㅓ'),
32
+ ('ㅖ', 'ㅣㅔ'),
33
+ ('ㅛ', 'ㅣㅗ'),
34
+ ('ㅠ', 'ㅣㅜ')
35
+ ]]
36
+
37
+ # List of (Latin alphabet, hangul) pairs:
38
+ _latin_to_hangul = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
39
+ ('a', '에이'),
40
+ ('b', '비'),
41
+ ('c', '시'),
42
+ ('d', '디'),
43
+ ('e', '이'),
44
+ ('f', '에프'),
45
+ ('g', '지'),
46
+ ('h', '에이치'),
47
+ ('i', '아이'),
48
+ ('j', '제이'),
49
+ ('k', '케이'),
50
+ ('l', '엘'),
51
+ ('m', '엠'),
52
+ ('n', '엔'),
53
+ ('o', '오'),
54
+ ('p', '피'),
55
+ ('q', '큐'),
56
+ ('r', '아르'),
57
+ ('s', '에스'),
58
+ ('t', '티'),
59
+ ('u', '유'),
60
+ ('v', '브이'),
61
+ ('w', '더블유'),
62
+ ('x', '엑스'),
63
+ ('y', '와이'),
64
+ ('z', '제트')
65
+ ]]
66
+
67
+ # List of (ipa, lazy ipa) pairs:
68
+ _ipa_to_lazy_ipa = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
69
+ ('t͡ɕ','ʧ'),
70
+ ('d͡ʑ','ʥ'),
71
+ ('ɲ','n^'),
72
+ ('ɕ','ʃ'),
73
+ ('ʷ','w'),
74
+ ('ɭ','l`'),
75
+ ('ʎ','ɾ'),
76
+ ('ɣ','ŋ'),
77
+ ('ɰ','ɯ'),
78
+ ('ʝ','j'),
79
+ ('ʌ','ə'),
80
+ ('ɡ','g'),
81
+ ('\u031a','#'),
82
+ ('\u0348','='),
83
+ ('\u031e',''),
84
+ ('\u0320',''),
85
+ ('\u0339','')
86
+ ]]
87
+
88
+
89
+ def latin_to_hangul(text):
90
+ for regex, replacement in _latin_to_hangul:
91
+ text = re.sub(regex, replacement, text)
92
+ return text
93
+
94
+
95
+ def divide_hangul(text):
96
+ text = j2hcj(h2j(text))
97
+ for regex, replacement in _hangul_divided:
98
+ text = re.sub(regex, replacement, text)
99
+ return text
100
+
101
+
102
+ def hangul_number(num, sino=True):
103
+ '''Reference https://github.com/Kyubyong/g2pK'''
104
+ num = re.sub(',', '', num)
105
+
106
+ if num == '0':
107
+ return '영'
108
+ if not sino and num == '20':
109
+ return '스무'
110
+
111
+ digits = '123456789'
112
+ names = '일이삼사오육칠팔구'
113
+ digit2name = {d: n for d, n in zip(digits, names)}
114
+
115
+ modifiers = '한 두 세 네 다섯 여섯 일곱 여덟 아홉'
116
+ decimals = '열 스물 서른 마흔 쉰 예순 일흔 여든 아흔'
117
+ digit2mod = {d: mod for d, mod in zip(digits, modifiers.split())}
118
+ digit2dec = {d: dec for d, dec in zip(digits, decimals.split())}
119
+
120
+ spelledout = []
121
+ for i, digit in enumerate(num):
122
+ i = len(num) - i - 1
123
+ if sino:
124
+ if i == 0:
125
+ name = digit2name.get(digit, '')
126
+ elif i == 1:
127
+ name = digit2name.get(digit, '') + '십'
128
+ name = name.replace('일십', '십')
129
+ else:
130
+ if i == 0:
131
+ name = digit2mod.get(digit, '')
132
+ elif i == 1:
133
+ name = digit2dec.get(digit, '')
134
+ if digit == '0':
135
+ if i % 4 == 0:
136
+ last_three = spelledout[-min(3, len(spelledout)):]
137
+ if ''.join(last_three) == '':
138
+ spelledout.append('')
139
+ continue
140
+ else:
141
+ spelledout.append('')
142
+ continue
143
+ if i == 2:
144
+ name = digit2name.get(digit, '') + '백'
145
+ name = name.replace('일백', '백')
146
+ elif i == 3:
147
+ name = digit2name.get(digit, '') + '천'
148
+ name = name.replace('일천', '천')
149
+ elif i == 4:
150
+ name = digit2name.get(digit, '') + '만'
151
+ name = name.replace('일만', '만')
152
+ elif i == 5:
153
+ name = digit2name.get(digit, '') + '십'
154
+ name = name.replace('일십', '십')
155
+ elif i == 6:
156
+ name = digit2name.get(digit, '') + '백'
157
+ name = name.replace('일백', '백')
158
+ elif i == 7:
159
+ name = digit2name.get(digit, '') + '천'
160
+ name = name.replace('일천', '천')
161
+ elif i == 8:
162
+ name = digit2name.get(digit, '') + '억'
163
+ elif i == 9:
164
+ name = digit2name.get(digit, '') + '십'
165
+ elif i == 10:
166
+ name = digit2name.get(digit, '') + '백'
167
+ elif i == 11:
168
+ name = digit2name.get(digit, '') + '천'
169
+ elif i == 12:
170
+ name = digit2name.get(digit, '') + '조'
171
+ elif i == 13:
172
+ name = digit2name.get(digit, '') + '십'
173
+ elif i == 14:
174
+ name = digit2name.get(digit, '') + '백'
175
+ elif i == 15:
176
+ name = digit2name.get(digit, '') + '천'
177
+ spelledout.append(name)
178
+ return ''.join(elem for elem in spelledout)
179
+
180
+
181
+ def number_to_hangul(text):
182
+ '''Reference https://github.com/Kyubyong/g2pK'''
183
+ tokens = set(re.findall(r'(\d[\d,]*)([\uac00-\ud71f]+)', text))
184
+ for token in tokens:
185
+ num, classifier = token
186
+ if classifier[:2] in _korean_classifiers or classifier[0] in _korean_classifiers:
187
+ spelledout = hangul_number(num, sino=False)
188
+ else:
189
+ spelledout = hangul_number(num, sino=True)
190
+ text = text.replace(f'{num}{classifier}', f'{spelledout}{classifier}')
191
+ # digit by digit for remaining digits
192
+ digits = '0123456789'
193
+ names = '영일이삼사오육칠팔구'
194
+ for d, n in zip(digits, names):
195
+ text = text.replace(d, n)
196
+ return text
197
+
198
+
199
+ def korean_to_lazy_ipa(text):
200
+ text = latin_to_hangul(text)
201
+ text = number_to_hangul(text)
202
+ text=re.sub('[\uac00-\ud7af]+',lambda x:ko_pron.romanise(x.group(0),'ipa').split('] ~ [')[0],text)
203
+ for regex, replacement in _ipa_to_lazy_ipa:
204
+ text = re.sub(regex, replacement, text)
205
+ return text
206
+
207
+
208
+ def korean_to_ipa(text):
209
+ text = korean_to_lazy_ipa(text)
210
+ return text.replace('ʧ','tʃ').replace('ʥ','dʑ')
text/mandarin.py ADDED
@@ -0,0 +1,330 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import sys
3
+ import re
4
+ from pypinyin import lazy_pinyin, BOPOMOFO
5
+ import jieba
6
+ import cn2an
7
+ import logging
8
+
9
+ logging.getLogger('jieba').setLevel(logging.WARNING)
10
+ jieba.set_dictionary(os.path.dirname(os.path.realpath(sys.argv[0]))+'/jieba/dict.txt')
11
+ jieba.initialize()
12
+
13
+
14
+ # List of (Latin alphabet, bopomofo) pairs:
15
+ _latin_to_bopomofo = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
16
+ ('a', 'ㄟˉ'),
17
+ ('b', 'ㄅㄧˋ'),
18
+ ('c', 'ㄙㄧˉ'),
19
+ ('d', 'ㄉㄧˋ'),
20
+ ('e', 'ㄧˋ'),
21
+ ('f', 'ㄝˊㄈㄨˋ'),
22
+ ('g', 'ㄐㄧˋ'),
23
+ ('h', 'ㄝˇㄑㄩˋ'),
24
+ ('i', 'ㄞˋ'),
25
+ ('j', 'ㄐㄟˋ'),
26
+ ('k', 'ㄎㄟˋ'),
27
+ ('l', 'ㄝˊㄛˋ'),
28
+ ('m', 'ㄝˊㄇㄨˋ'),
29
+ ('n', 'ㄣˉ'),
30
+ ('o', 'ㄡˉ'),
31
+ ('p', 'ㄆㄧˉ'),
32
+ ('q', 'ㄎㄧㄡˉ'),
33
+ ('r', 'ㄚˋ'),
34
+ ('s', 'ㄝˊㄙˋ'),
35
+ ('t', 'ㄊㄧˋ'),
36
+ ('u', 'ㄧㄡˉ'),
37
+ ('v', 'ㄨㄧˉ'),
38
+ ('w', 'ㄉㄚˋㄅㄨˋㄌㄧㄡˋ'),
39
+ ('x', 'ㄝˉㄎㄨˋㄙˋ'),
40
+ ('y', 'ㄨㄞˋ'),
41
+ ('z', 'ㄗㄟˋ')
42
+ ]]
43
+
44
+ # List of (bopomofo, romaji) pairs:
45
+ _bopomofo_to_romaji = [(re.compile('%s' % x[0]), x[1]) for x in [
46
+ ('ㄅㄛ', 'p⁼wo'),
47
+ ('ㄆㄛ', 'pʰwo'),
48
+ ('ㄇㄛ', 'mwo'),
49
+ ('ㄈㄛ', 'fwo'),
50
+ ('ㄅ', 'p⁼'),
51
+ ('ㄆ', 'pʰ'),
52
+ ('ㄇ', 'm'),
53
+ ('ㄈ', 'f'),
54
+ ('ㄉ', 't⁼'),
55
+ ('ㄊ', 'tʰ'),
56
+ ('ㄋ', 'n'),
57
+ ('ㄌ', 'l'),
58
+ ('ㄍ', 'k⁼'),
59
+ ('ㄎ', 'kʰ'),
60
+ ('ㄏ', 'h'),
61
+ ('ㄐ', 'ʧ⁼'),
62
+ ('ㄑ', 'ʧʰ'),
63
+ ('ㄒ', 'ʃ'),
64
+ ('ㄓ', 'ʦ`⁼'),
65
+ ('ㄔ', 'ʦ`ʰ'),
66
+ ('ㄕ', 's`'),
67
+ ('ㄖ', 'ɹ`'),
68
+ ('ㄗ', 'ʦ⁼'),
69
+ ('ㄘ', 'ʦʰ'),
70
+ ('ㄙ', 's'),
71
+ ('ㄚ', 'a'),
72
+ ('ㄛ', 'o'),
73
+ ('ㄜ', 'ə'),
74
+ ('ㄝ', 'e'),
75
+ ('ㄞ', 'ai'),
76
+ ('ㄟ', 'ei'),
77
+ ('ㄠ', 'au'),
78
+ ('ㄡ', 'ou'),
79
+ ('ㄧㄢ', 'yeNN'),
80
+ ('ㄢ', 'aNN'),
81
+ ('ㄧㄣ', 'iNN'),
82
+ ('ㄣ', 'əNN'),
83
+ ('ㄤ', 'aNg'),
84
+ ('ㄧㄥ', 'iNg'),
85
+ ('ㄨㄥ', 'uNg'),
86
+ ('ㄩㄥ', 'yuNg'),
87
+ ('ㄥ', 'əNg'),
88
+ ('ㄦ', 'əɻ'),
89
+ ('ㄧ', 'i'),
90
+ ('ㄨ', 'u'),
91
+ ('ㄩ', 'ɥ'),
92
+ ('ˉ', '→'),
93
+ ('ˊ', '↑'),
94
+ ('ˇ', '↓↑'),
95
+ ('ˋ', '↓'),
96
+ ('˙', ''),
97
+ (',', ','),
98
+ ('。', '.'),
99
+ ('!', '!'),
100
+ ('?', '?'),
101
+ ('—', '-')
102
+ ]]
103
+
104
+ # List of (romaji, ipa) pairs:
105
+ _romaji_to_ipa = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
106
+ ('ʃy', 'ʃ'),
107
+ ('ʧʰy', 'ʧʰ'),
108
+ ('ʧ⁼y', 'ʧ⁼'),
109
+ ('NN', 'n'),
110
+ ('Ng', 'ŋ'),
111
+ ('y', 'j'),
112
+ ('h', 'x')
113
+ ]]
114
+
115
+ # List of (bopomofo, ipa) pairs:
116
+ _bopomofo_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
117
+ ('ㄅㄛ', 'p⁼wo'),
118
+ ('ㄆㄛ', 'pʰwo'),
119
+ ('ㄇㄛ', 'mwo'),
120
+ ('ㄈㄛ', 'fwo'),
121
+ ('ㄅ', 'p⁼'),
122
+ ('ㄆ', 'pʰ'),
123
+ ('ㄇ', 'm'),
124
+ ('ㄈ', 'f'),
125
+ ('ㄉ', 't⁼'),
126
+ ('ㄊ', 'tʰ'),
127
+ ('ㄋ', 'n'),
128
+ ('ㄌ', 'l'),
129
+ ('ㄍ', 'k⁼'),
130
+ ('ㄎ', 'kʰ'),
131
+ ('ㄏ', 'x'),
132
+ ('ㄐ', 'tʃ⁼'),
133
+ ('ㄑ', 'tʃʰ'),
134
+ ('ㄒ', 'ʃ'),
135
+ ('ㄓ', 'ts`⁼'),
136
+ ('ㄔ', 'ts`ʰ'),
137
+ ('ㄕ', 's`'),
138
+ ('ㄖ', 'ɹ`'),
139
+ ('ㄗ', 'ts⁼'),
140
+ ('ㄘ', 'tsʰ'),
141
+ ('ㄙ', 's'),
142
+ ('ㄚ', 'a'),
143
+ ('ㄛ', 'o'),
144
+ ('ㄜ', 'ə'),
145
+ ('ㄝ', 'ɛ'),
146
+ ('ㄞ', 'aɪ'),
147
+ ('ㄟ', 'eɪ'),
148
+ ('ㄠ', 'ɑʊ'),
149
+ ('ㄡ', 'oʊ'),
150
+ ('ㄧㄢ', 'jɛn'),
151
+ ('ㄩㄢ', 'ɥæn'),
152
+ ('ㄢ', 'an'),
153
+ ('ㄧㄣ', 'in'),
154
+ ('ㄩㄣ', 'ɥn'),
155
+ ('ㄣ', 'ən'),
156
+ ('ㄤ', 'ɑŋ'),
157
+ ('ㄧㄥ', 'iŋ'),
158
+ ('ㄨㄥ', 'ʊŋ'),
159
+ ('ㄩㄥ', 'jʊŋ'),
160
+ ('ㄥ', 'əŋ'),
161
+ ('ㄦ', 'əɻ'),
162
+ ('ㄧ', 'i'),
163
+ ('ㄨ', 'u'),
164
+ ('ㄩ', 'ɥ'),
165
+ ('ˉ', '→'),
166
+ ('ˊ', '↑'),
167
+ ('ˇ', '↓↑'),
168
+ ('ˋ', '↓'),
169
+ ('˙', ''),
170
+ (',', ','),
171
+ ('。', '.'),
172
+ ('!', '!'),
173
+ ('?', '?'),
174
+ ('—', '-')
175
+ ]]
176
+
177
+ # List of (bopomofo, ipa2) pairs:
178
+ _bopomofo_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
179
+ ('ㄅㄛ', 'pwo'),
180
+ ('ㄆㄛ', 'pʰwo'),
181
+ ('ㄇㄛ', 'mwo'),
182
+ ('ㄈㄛ', 'fwo'),
183
+ ('ㄅ', 'p'),
184
+ ('ㄆ', 'pʰ'),
185
+ ('ㄇ', 'm'),
186
+ ('ㄈ', 'f'),
187
+ ('ㄉ', 't'),
188
+ ('ㄊ', 'tʰ'),
189
+ ('ㄋ', 'n'),
190
+ ('ㄌ', 'l'),
191
+ ('ㄍ', 'k'),
192
+ ('ㄎ', 'kʰ'),
193
+ ('ㄏ', 'h'),
194
+ ('ㄐ', 'tɕ'),
195
+ ('ㄑ', 'tɕʰ'),
196
+ ('ㄒ', 'ɕ'),
197
+ ('ㄓ', 'tʂ'),
198
+ ('ㄔ', 'tʂʰ'),
199
+ ('ㄕ', 'ʂ'),
200
+ ('ㄖ', 'ɻ'),
201
+ ('ㄗ', 'ts'),
202
+ ('ㄘ', 'tsʰ'),
203
+ ('ㄙ', 's'),
204
+ ('ㄚ', 'a'),
205
+ ('ㄛ', 'o'),
206
+ ('ㄜ', 'ɤ'),
207
+ ('ㄝ', 'ɛ'),
208
+ ('ㄞ', 'aɪ'),
209
+ ('ㄟ', 'eɪ'),
210
+ ('ㄠ', 'ɑʊ'),
211
+ ('ㄡ', 'oʊ'),
212
+ ('ㄧㄢ', 'jɛn'),
213
+ ('ㄩㄢ', 'yæn'),
214
+ ('ㄢ', 'an'),
215
+ ('ㄧㄣ', 'in'),
216
+ ('ㄩㄣ', 'yn'),
217
+ ('ㄣ', 'ən'),
218
+ ('ㄤ', 'ɑŋ'),
219
+ ('ㄧㄥ', 'iŋ'),
220
+ ('ㄨㄥ', 'ʊŋ'),
221
+ ('ㄩㄥ', 'jʊŋ'),
222
+ ('ㄥ', 'ɤŋ'),
223
+ ('ㄦ', 'əɻ'),
224
+ ('ㄧ', 'i'),
225
+ ('ㄨ', 'u'),
226
+ ('ㄩ', 'y'),
227
+ ('ˉ', '˥'),
228
+ ('ˊ', '˧˥'),
229
+ ('ˇ', '˨˩˦'),
230
+ ('ˋ', '˥˩'),
231
+ ('˙', ''),
232
+ (',', ','),
233
+ ('。', '.'),
234
+ ('!', '!'),
235
+ ('?', '?'),
236
+ ('—', '-')
237
+ ]]
238
+
239
+
240
+ def number_to_chinese(text):
241
+ numbers = re.findall(r'\d+(?:\.?\d+)?', text)
242
+ for number in numbers:
243
+ text = text.replace(number, cn2an.an2cn(number), 1)
244
+ return text
245
+
246
+
247
+ def chinese_to_bopomofo(text):
248
+ text = text.replace('、', ',').replace(';', ',').replace(':', ',')
249
+ words = jieba.lcut(text, cut_all=False)
250
+ text = ''
251
+ for word in words:
252
+ bopomofos = lazy_pinyin(word, BOPOMOFO)
253
+ if not re.search('[\u4e00-\u9fff]', word):
254
+ text += word
255
+ continue
256
+ for i in range(len(bopomofos)):
257
+ bopomofos[i] = re.sub(r'([\u3105-\u3129])$', r'\1ˉ', bopomofos[i])
258
+ if text != '':
259
+ text += ' '
260
+ text += ''.join(bopomofos)
261
+ return text
262
+
263
+
264
+ def latin_to_bopomofo(text):
265
+ for regex, replacement in _latin_to_bopomofo:
266
+ text = re.sub(regex, replacement, text)
267
+ return text
268
+
269
+
270
+ def bopomofo_to_romaji(text):
271
+ for regex, replacement in _bopomofo_to_romaji:
272
+ text = re.sub(regex, replacement, text)
273
+ return text
274
+
275
+
276
+ def bopomofo_to_ipa(text):
277
+ for regex, replacement in _bopomofo_to_ipa:
278
+ text = re.sub(regex, replacement, text)
279
+ return text
280
+
281
+
282
+ def bopomofo_to_ipa2(text):
283
+ for regex, replacement in _bopomofo_to_ipa2:
284
+ text = re.sub(regex, replacement, text)
285
+ return text
286
+
287
+
288
+ def chinese_to_romaji(text):
289
+ text = number_to_chinese(text)
290
+ text = chinese_to_bopomofo(text)
291
+ text = latin_to_bopomofo(text)
292
+ text = bopomofo_to_romaji(text)
293
+ text = re.sub('i([aoe])', r'y\1', text)
294
+ text = re.sub('u([aoəe])', r'w\1', text)
295
+ text = re.sub('([ʦsɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
296
+ r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
297
+ text = re.sub('([ʦs][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
298
+ return text
299
+
300
+
301
+ def chinese_to_lazy_ipa(text):
302
+ text = chinese_to_romaji(text)
303
+ for regex, replacement in _romaji_to_ipa:
304
+ text = re.sub(regex, replacement, text)
305
+ return text
306
+
307
+
308
+ def chinese_to_ipa(text):
309
+ text = number_to_chinese(text)
310
+ text = chinese_to_bopomofo(text)
311
+ text = latin_to_bopomofo(text)
312
+ text = bopomofo_to_ipa(text)
313
+ text = re.sub('i([aoe])', r'j\1', text)
314
+ text = re.sub('u([aoəe])', r'w\1', text)
315
+ text = re.sub('([sɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
316
+ r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
317
+ text = re.sub('([s][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
318
+ return text
319
+
320
+
321
+ def chinese_to_ipa2(text):
322
+ text = number_to_chinese(text)
323
+ text = chinese_to_bopomofo(text)
324
+ text = latin_to_bopomofo(text)
325
+ text = bopomofo_to_ipa2(text)
326
+ text = re.sub(r'i([aoe])', r'j\1', text)
327
+ text = re.sub(r'u([aoəe])', r'w\1', text)
328
+ text = re.sub(r'([ʂɹ]ʰ?)([˩˨˧˦˥ ]+|$)', r'\1ʅ\2', text)
329
+ text = re.sub(r'(sʰ?)([˩˨˧˦˥ ]+|$)', r'\1ɿ\2', text)
330
+ return text
text/ngu_dialect.py ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ import opencc
3
+
4
+
5
+ dialects = {'SZ': 'suzhou', 'WX': 'wuxi', 'CZ': 'changzhou', 'HZ': 'hangzhou',
6
+ 'SX': 'shaoxing', 'NB': 'ningbo', 'JJ': 'jingjiang', 'YX': 'yixing',
7
+ 'JD': 'jiading', 'ZR': 'zhenru', 'PH': 'pinghu', 'TX': 'tongxiang',
8
+ 'JS': 'jiashan', 'HN': 'xiashi', 'LP': 'linping', 'XS': 'xiaoshan',
9
+ 'FY': 'fuyang', 'RA': 'ruao', 'CX': 'cixi', 'SM': 'sanmen',
10
+ 'TT': 'tiantai', 'WZ': 'wenzhou', 'SC': 'suichang', 'YB': 'youbu'}
11
+
12
+ converters = {}
13
+
14
+ for dialect in dialects.values():
15
+ try:
16
+ converters[dialect] = opencc.OpenCC(dialect)
17
+ except:
18
+ pass
19
+
20
+
21
+ def ngu_dialect_to_ipa(text, dialect):
22
+ dialect = dialects[dialect]
23
+ text = converters[dialect].convert(text).replace('-','').replace('$',' ')
24
+ text = re.sub(r'[、;:]', ',', text)
25
+ text = re.sub(r'\s*,\s*', ', ', text)
26
+ text = re.sub(r'\s*。\s*', '. ', text)
27
+ text = re.sub(r'\s*?\s*', '? ', text)
28
+ text = re.sub(r'\s*!\s*', '! ', text)
29
+ text = re.sub(r'\s*$', '', text)
30
+ return text
text/sanskrit.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from indic_transliteration import sanscript
3
+
4
+
5
+ # List of (iast, ipa) pairs:
6
+ _iast_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
7
+ ('a', 'ə'),
8
+ ('ā', 'aː'),
9
+ ('ī', 'iː'),
10
+ ('ū', 'uː'),
11
+ ('ṛ', 'ɹ`'),
12
+ ('ṝ', 'ɹ`ː'),
13
+ ('ḷ', 'l`'),
14
+ ('ḹ', 'l`ː'),
15
+ ('e', 'eː'),
16
+ ('o', 'oː'),
17
+ ('k', 'k⁼'),
18
+ ('k⁼h', 'kʰ'),
19
+ ('g', 'g⁼'),
20
+ ('g⁼h', 'gʰ'),
21
+ ('ṅ', 'ŋ'),
22
+ ('c', 'ʧ⁼'),
23
+ ('ʧ⁼h', 'ʧʰ'),
24
+ ('j', 'ʥ⁼'),
25
+ ('ʥ⁼h', 'ʥʰ'),
26
+ ('ñ', 'n^'),
27
+ ('ṭ', 't`⁼'),
28
+ ('t`⁼h', 't`ʰ'),
29
+ ('ḍ', 'd`⁼'),
30
+ ('d`⁼h', 'd`ʰ'),
31
+ ('ṇ', 'n`'),
32
+ ('t', 't⁼'),
33
+ ('t⁼h', 'tʰ'),
34
+ ('d', 'd⁼'),
35
+ ('d⁼h', 'dʰ'),
36
+ ('p', 'p⁼'),
37
+ ('p⁼h', 'pʰ'),
38
+ ('b', 'b⁼'),
39
+ ('b⁼h', 'bʰ'),
40
+ ('y', 'j'),
41
+ ('ś', 'ʃ'),
42
+ ('ṣ', 's`'),
43
+ ('r', 'ɾ'),
44
+ ('l̤', 'l`'),
45
+ ('h', 'ɦ'),
46
+ ("'", ''),
47
+ ('~', '^'),
48
+ ('ṃ', '^')
49
+ ]]
50
+
51
+
52
+ def devanagari_to_ipa(text):
53
+ text = text.replace('ॐ', 'ओम्')
54
+ text = re.sub(r'\s*।\s*$', '.', text)
55
+ text = re.sub(r'\s*।\s*', ', ', text)
56
+ text = re.sub(r'\s*॥', '.', text)
57
+ text = sanscript.transliterate(text, sanscript.DEVANAGARI, sanscript.IAST)
58
+ for regex, replacement in _iast_to_ipa:
59
+ text = re.sub(regex, replacement, text)
60
+ text = re.sub('(.)[`ː]*ḥ', lambda x: x.group(0)
61
+ [:-1]+'h'+x.group(1)+'*', text)
62
+ return text
text/shanghainese.py ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ import cn2an
3
+ import opencc
4
+
5
+
6
+ converter = opencc.OpenCC('zaonhe')
7
+
8
+ # List of (Latin alphabet, ipa) pairs:
9
+ _latin_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
10
+ ('A', 'ᴇ'),
11
+ ('B', 'bi'),
12
+ ('C', 'si'),
13
+ ('D', 'di'),
14
+ ('E', 'i'),
15
+ ('F', 'ᴇf'),
16
+ ('G', 'dʑi'),
17
+ ('H', 'ᴇtɕʰ'),
18
+ ('I', 'ᴀi'),
19
+ ('J', 'dʑᴇ'),
20
+ ('K', 'kʰᴇ'),
21
+ ('L', 'ᴇl'),
22
+ ('M', 'ᴇm'),
23
+ ('N', 'ᴇn'),
24
+ ('O', 'o'),
25
+ ('P', 'pʰi'),
26
+ ('Q', 'kʰiu'),
27
+ ('R', 'ᴀl'),
28
+ ('S', 'ᴇs'),
29
+ ('T', 'tʰi'),
30
+ ('U', 'ɦiu'),
31
+ ('V', 'vi'),
32
+ ('W', 'dᴀbɤliu'),
33
+ ('X', 'ᴇks'),
34
+ ('Y', 'uᴀi'),
35
+ ('Z', 'zᴇ')
36
+ ]]
37
+
38
+
39
+ def _number_to_shanghainese(num):
40
+ num = cn2an.an2cn(num).replace('一十','十').replace('二十', '廿').replace('二', '两')
41
+ return re.sub(r'((?:^|[^三四五六七八九])十|廿)两', r'\1二', num)
42
+
43
+
44
+ def number_to_shanghainese(text):
45
+ return re.sub(r'\d+(?:\.?\d+)?', lambda x: _number_to_shanghainese(x.group()), text)
46
+
47
+
48
+ def latin_to_ipa(text):
49
+ for regex, replacement in _latin_to_ipa:
50
+ text = re.sub(regex, replacement, text)
51
+ return text
52
+
53
+
54
+ def shanghainese_to_ipa(text):
55
+ text = number_to_shanghainese(text.upper())
56
+ text = converter.convert(text).replace('-','').replace('$',' ')
57
+ text = re.sub(r'[A-Z]', lambda x: latin_to_ipa(x.group())+' ', text)
58
+ text = re.sub(r'[、;:]', ',', text)
59
+ text = re.sub(r'\s*,\s*', ', ', text)
60
+ text = re.sub(r'\s*。\s*', '. ', text)
61
+ text = re.sub(r'\s*?\s*', '? ', text)
62
+ text = re.sub(r'\s*!\s*', '! ', text)
63
+ text = re.sub(r'\s*$', '', text)
64
+ return text
text/thai.py ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from num_thai.thainumbers import NumThai
3
+
4
+
5
+ num = NumThai()
6
+
7
+ # List of (Latin alphabet, Thai) pairs:
8
+ _latin_to_thai = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
9
+ ('a', 'เอ'),
10
+ ('b','บี'),
11
+ ('c','ซี'),
12
+ ('d','ดี'),
13
+ ('e','อี'),
14
+ ('f','เอฟ'),
15
+ ('g','จี'),
16
+ ('h','เอช'),
17
+ ('i','ไอ'),
18
+ ('j','เจ'),
19
+ ('k','เค'),
20
+ ('l','แอล'),
21
+ ('m','เอ็ม'),
22
+ ('n','เอ็น'),
23
+ ('o','โอ'),
24
+ ('p','พี'),
25
+ ('q','คิว'),
26
+ ('r','แอร์'),
27
+ ('s','เอส'),
28
+ ('t','ที'),
29
+ ('u','ยู'),
30
+ ('v','วี'),
31
+ ('w','ดับเบิลยู'),
32
+ ('x','เอ็กซ์'),
33
+ ('y','วาย'),
34
+ ('z','ซี')
35
+ ]]
36
+
37
+
38
+ def num_to_thai(text):
39
+ return re.sub(r'(?:\d+(?:,?\d+)?)+(?:\.\d+(?:,?\d+)?)?', lambda x: ''.join(num.NumberToTextThai(float(x.group(0).replace(',', '')))), text)
40
+
41
+ def latin_to_thai(text):
42
+ for regex, replacement in _latin_to_thai:
43
+ text = re.sub(regex, replacement, text)
44
+ return text
transforms.py ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+
4
+ import numpy as np
5
+
6
+
7
+ DEFAULT_MIN_BIN_WIDTH = 1e-3
8
+ DEFAULT_MIN_BIN_HEIGHT = 1e-3
9
+ DEFAULT_MIN_DERIVATIVE = 1e-3
10
+
11
+
12
+ def piecewise_rational_quadratic_transform(inputs,
13
+ unnormalized_widths,
14
+ unnormalized_heights,
15
+ unnormalized_derivatives,
16
+ inverse=False,
17
+ tails=None,
18
+ tail_bound=1.,
19
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
20
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
21
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
22
+
23
+ if tails is None:
24
+ spline_fn = rational_quadratic_spline
25
+ spline_kwargs = {}
26
+ else:
27
+ spline_fn = unconstrained_rational_quadratic_spline
28
+ spline_kwargs = {
29
+ 'tails': tails,
30
+ 'tail_bound': tail_bound
31
+ }
32
+
33
+ outputs, logabsdet = spline_fn(
34
+ inputs=inputs,
35
+ unnormalized_widths=unnormalized_widths,
36
+ unnormalized_heights=unnormalized_heights,
37
+ unnormalized_derivatives=unnormalized_derivatives,
38
+ inverse=inverse,
39
+ min_bin_width=min_bin_width,
40
+ min_bin_height=min_bin_height,
41
+ min_derivative=min_derivative,
42
+ **spline_kwargs
43
+ )
44
+ return outputs, logabsdet
45
+
46
+
47
+ def searchsorted(bin_locations, inputs, eps=1e-6):
48
+ bin_locations[..., -1] += eps
49
+ return torch.sum(
50
+ inputs[..., None] >= bin_locations,
51
+ dim=-1
52
+ ) - 1
53
+
54
+
55
+ def unconstrained_rational_quadratic_spline(inputs,
56
+ unnormalized_widths,
57
+ unnormalized_heights,
58
+ unnormalized_derivatives,
59
+ inverse=False,
60
+ tails='linear',
61
+ tail_bound=1.,
62
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
63
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
64
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
65
+ inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
66
+ outside_interval_mask = ~inside_interval_mask
67
+
68
+ outputs = torch.zeros_like(inputs)
69
+ logabsdet = torch.zeros_like(inputs)
70
+
71
+ if tails == 'linear':
72
+ unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
73
+ constant = np.log(np.exp(1 - min_derivative) - 1)
74
+ unnormalized_derivatives[..., 0] = constant
75
+ unnormalized_derivatives[..., -1] = constant
76
+
77
+ outputs[outside_interval_mask] = inputs[outside_interval_mask]
78
+ logabsdet[outside_interval_mask] = 0
79
+ else:
80
+ raise RuntimeError('{} tails are not implemented.'.format(tails))
81
+
82
+ outputs[inside_interval_mask], logabsdet[inside_interval_mask] = rational_quadratic_spline(
83
+ inputs=inputs[inside_interval_mask],
84
+ unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
85
+ unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
86
+ unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
87
+ inverse=inverse,
88
+ left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound,
89
+ min_bin_width=min_bin_width,
90
+ min_bin_height=min_bin_height,
91
+ min_derivative=min_derivative
92
+ )
93
+
94
+ return outputs, logabsdet
95
+
96
+ def rational_quadratic_spline(inputs,
97
+ unnormalized_widths,
98
+ unnormalized_heights,
99
+ unnormalized_derivatives,
100
+ inverse=False,
101
+ left=0., right=1., bottom=0., top=1.,
102
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
103
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
104
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
105
+ if torch.min(inputs) < left or torch.max(inputs) > right:
106
+ raise ValueError('Input to a transform is not within its domain')
107
+
108
+ num_bins = unnormalized_widths.shape[-1]
109
+
110
+ if min_bin_width * num_bins > 1.0:
111
+ raise ValueError('Minimal bin width too large for the number of bins')
112
+ if min_bin_height * num_bins > 1.0:
113
+ raise ValueError('Minimal bin height too large for the number of bins')
114
+
115
+ widths = F.softmax(unnormalized_widths, dim=-1)
116
+ widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
117
+ cumwidths = torch.cumsum(widths, dim=-1)
118
+ cumwidths = F.pad(cumwidths, pad=(1, 0), mode='constant', value=0.0)
119
+ cumwidths = (right - left) * cumwidths + left
120
+ cumwidths[..., 0] = left
121
+ cumwidths[..., -1] = right
122
+ widths = cumwidths[..., 1:] - cumwidths[..., :-1]
123
+
124
+ derivatives = min_derivative + F.softplus(unnormalized_derivatives)
125
+
126
+ heights = F.softmax(unnormalized_heights, dim=-1)
127
+ heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
128
+ cumheights = torch.cumsum(heights, dim=-1)
129
+ cumheights = F.pad(cumheights, pad=(1, 0), mode='constant', value=0.0)
130
+ cumheights = (top - bottom) * cumheights + bottom
131
+ cumheights[..., 0] = bottom
132
+ cumheights[..., -1] = top
133
+ heights = cumheights[..., 1:] - cumheights[..., :-1]
134
+
135
+ if inverse:
136
+ bin_idx = searchsorted(cumheights, inputs)[..., None]
137
+ else:
138
+ bin_idx = searchsorted(cumwidths, inputs)[..., None]
139
+
140
+ input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
141
+ input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
142
+
143
+ input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
144
+ delta = heights / widths
145
+ input_delta = delta.gather(-1, bin_idx)[..., 0]
146
+
147
+ input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
148
+ input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
149
+
150
+ input_heights = heights.gather(-1, bin_idx)[..., 0]
151
+
152
+ if inverse:
153
+ a = (((inputs - input_cumheights) * (input_derivatives
154
+ + input_derivatives_plus_one
155
+ - 2 * input_delta)
156
+ + input_heights * (input_delta - input_derivatives)))
157
+ b = (input_heights * input_derivatives
158
+ - (inputs - input_cumheights) * (input_derivatives
159
+ + input_derivatives_plus_one
160
+ - 2 * input_delta))
161
+ c = - input_delta * (inputs - input_cumheights)
162
+
163
+ discriminant = b.pow(2) - 4 * a * c
164
+ assert (discriminant >= 0).all()
165
+
166
+ root = (2 * c) / (-b - torch.sqrt(discriminant))
167
+ outputs = root * input_bin_widths + input_cumwidths
168
+
169
+ theta_one_minus_theta = root * (1 - root)
170
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
171
+ * theta_one_minus_theta)
172
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * root.pow(2)
173
+ + 2 * input_delta * theta_one_minus_theta
174
+ + input_derivatives * (1 - root).pow(2))
175
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
176
+
177
+ return outputs, -logabsdet
178
+ else:
179
+ theta = (inputs - input_cumwidths) / input_bin_widths
180
+ theta_one_minus_theta = theta * (1 - theta)
181
+
182
+ numerator = input_heights * (input_delta * theta.pow(2)
183
+ + input_derivatives * theta_one_minus_theta)
184
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
185
+ * theta_one_minus_theta)
186
+ outputs = input_cumheights + numerator / denominator
187
+
188
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2)
189
+ + 2 * input_delta * theta_one_minus_theta
190
+ + input_derivatives * (1 - theta).pow(2))
191
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
192
+
193
+ return outputs, logabsdet
utils.py ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
+ import os
3
+ import sys
4
+ from io import BytesIO
5
+ from json import loads
6
+ import av
7
+ import pilk
8
+ from flask import Flask
9
+ from torch import load, FloatTensor
10
+ from numpy import float32
11
+ import librosa
12
+
13
+ from voice import Voice
14
+
15
+ app = Flask(__name__)
16
+ app.config.from_pyfile("config.py")
17
+
18
+
19
+ class HParams():
20
+ def __init__(self, **kwargs):
21
+ for k, v in kwargs.items():
22
+ if type(v) == dict:
23
+ v = HParams(**v)
24
+ self[k] = v
25
+
26
+ def keys(self):
27
+ return self.__dict__.keys()
28
+
29
+ def items(self):
30
+ return self.__dict__.items()
31
+
32
+ def values(self):
33
+ return self.__dict__.values()
34
+
35
+ def __len__(self):
36
+ return len(self.__dict__)
37
+
38
+ def __getitem__(self, key):
39
+ return getattr(self, key)
40
+
41
+ def __setitem__(self, key, value):
42
+ return setattr(self, key, value)
43
+
44
+ def __contains__(self, key):
45
+ return key in self.__dict__
46
+
47
+ def __repr__(self):
48
+ return self.__dict__.__repr__()
49
+
50
+
51
+ def load_checkpoint(checkpoint_path, model):
52
+ checkpoint_dict = load(checkpoint_path, map_location='cpu')
53
+ iteration = checkpoint_dict['iteration']
54
+ saved_state_dict = checkpoint_dict['model']
55
+ if hasattr(model, 'module'):
56
+ state_dict = model.module.state_dict()
57
+ else:
58
+ state_dict = model.state_dict()
59
+ new_state_dict = {}
60
+ for k, v in state_dict.items():
61
+ try:
62
+ new_state_dict[k] = saved_state_dict[k]
63
+ except:
64
+ logging.info("%s is not in the checkpoint" % k)
65
+ new_state_dict[k] = v
66
+ if hasattr(model, 'module'):
67
+ model.module.load_state_dict(new_state_dict)
68
+ else:
69
+ model.load_state_dict(new_state_dict)
70
+ logging.info("Loaded checkpoint '{}' (iteration {})".format(
71
+ checkpoint_path, iteration))
72
+ return
73
+
74
+
75
+ def get_hparams_from_file(config_path):
76
+ with open(config_path, "r") as f:
77
+ data = f.read()
78
+ config = loads(data)
79
+
80
+ hparams = HParams(**config)
81
+ return hparams
82
+
83
+
84
+ def load_audio_to_torch(full_path, target_sampling_rate):
85
+ audio, sampling_rate = librosa.load(full_path, sr=target_sampling_rate, mono=True)
86
+ return FloatTensor(audio.astype(float32))
87
+
88
+
89
+ def wav2ogg(input, output):
90
+ with av.open(input, 'rb') as i:
91
+ with av.open(output, 'wb', format='ogg') as o:
92
+ out_stream = o.add_stream('libvorbis')
93
+ for frame in i.decode(audio=0):
94
+ for p in out_stream.encode(frame):
95
+ o.mux(p)
96
+
97
+ for p in out_stream.encode(None):
98
+ o.mux(p)
99
+
100
+
101
+ # def wav2silk(input, output):
102
+ # with av.open(input) as in_wav:
103
+ # in_stream = in_wav.streams.audio[0]
104
+ # sample_rate = in_stream.codec_context.sample_rate
105
+ # with BytesIO() as pcm:
106
+ # with av.open(pcm, 'w', 's16le') as out_pcm:
107
+ # out_stream = out_pcm.add_stream(
108
+ # 'pcm_s16le',
109
+ # rate=sample_rate,
110
+ # layout='mono'
111
+ # )
112
+ # for frame in in_wav.decode(in_stream):
113
+ # frame.pts = None
114
+ # for packet in out_stream.encode(frame):
115
+ # out_pcm.mux(packet)
116
+ #
117
+ # pilk.encode(out_pcm, output, pcm_rate=sample_rate, tencent=True)
118
+
119
+
120
+ def to_pcm(in_path: str) -> tuple[str, int]:
121
+ out_path = os.path.splitext(in_path)[0] + '.pcm'
122
+ with av.open(in_path) as in_container:
123
+ in_stream = in_container.streams.audio[0]
124
+ sample_rate = in_stream.codec_context.sample_rate
125
+ with av.open(out_path, 'w', 's16le') as out_container:
126
+ out_stream = out_container.add_stream(
127
+ 'pcm_s16le',
128
+ rate=sample_rate,
129
+ layout='mono'
130
+ )
131
+ try:
132
+ for frame in in_container.decode(in_stream):
133
+ frame.pts = None
134
+ for packet in out_stream.encode(frame):
135
+ out_container.mux(packet)
136
+ except:
137
+ pass
138
+ return out_path, sample_rate
139
+
140
+
141
+ def convert_to_silk(media_path: str) -> str:
142
+ pcm_path, sample_rate = to_pcm(media_path)
143
+ silk_path = os.path.splitext(pcm_path)[0] + '.silk'
144
+ pilk.encode(pcm_path, silk_path, pcm_rate=sample_rate, tencent=True)
145
+ os.remove(pcm_path)
146
+ return silk_path
147
+
148
+
149
+ def clean_folder(folder_path):
150
+ for filename in os.listdir(folder_path):
151
+ file_path = os.path.join(folder_path, filename)
152
+ # 如果是文件,则删除文件
153
+ if os.path.isfile(file_path):
154
+ os.remove(file_path)
155
+
156
+
157
+ def merge_model(merging_model):
158
+ voice_obj = []
159
+ voice_speakers = []
160
+ new_id = 0
161
+ out_path = os.path.dirname(os.path.realpath(sys.argv[0])) + "/out_silk"
162
+ for obj_id, i in enumerate(merging_model):
163
+ obj = Voice(i[0], i[1], out_path)
164
+
165
+ for id, name in enumerate(obj.return_speakers()):
166
+ voice_obj.append([int(id), obj, obj_id])
167
+ voice_speakers.append({new_id: name})
168
+
169
+ new_id += 1
170
+
171
+ return voice_obj, voice_speakers
voice.py ADDED
@@ -0,0 +1,231 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ import librosa
4
+ from scipy.io.wavfile import write
5
+ from mel_processing import spectrogram_torch
6
+ from text import text_to_sequence, _clean_text
7
+ from models import SynthesizerTrn
8
+ import utils
9
+ import commons
10
+ import sys
11
+ import re
12
+ import numpy as np
13
+ # import torch
14
+ # torch.set_num_threads(1) #设置torch线程为1,防止多任务推理时服务崩溃,但flask仍然会使用多线程
15
+ from torch import no_grad, LongTensor, inference_mode, FloatTensor
16
+ import audonnx
17
+ import uuid
18
+ from io import BytesIO
19
+
20
+
21
+ class Voice:
22
+ def __init__(self, model, config, out_path=None):
23
+ self.out_path = out_path
24
+ if not os.path.exists(self.out_path):
25
+ try:
26
+ os.mkdir(self.out_path)
27
+ except:
28
+ pass
29
+
30
+ self.hps_ms = utils.get_hparams_from_file(config)
31
+ self.n_speakers = self.hps_ms.data.n_speakers if 'n_speakers' in self.hps_ms.data.keys() else 0
32
+ self.n_symbols = len(self.hps_ms.symbols) if 'symbols' in self.hps_ms.keys() else 0
33
+ self.speakers = self.hps_ms.speakers if 'speakers' in self.hps_ms.keys() else ['0']
34
+ self.use_f0 = self.hps_ms.data.use_f0 if 'use_f0' in self.hps_ms.data.keys() else False
35
+ self.emotion_embedding = self.hps_ms.data.emotion_embedding if 'emotion_embedding' in self.hps_ms.data.keys() else False
36
+
37
+ self.net_g_ms = SynthesizerTrn(
38
+ self.n_symbols,
39
+ self.hps_ms.data.filter_length // 2 + 1,
40
+ self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
41
+ n_speakers=self.n_speakers,
42
+ emotion_embedding=self.emotion_embedding,
43
+ **self.hps_ms.model)
44
+ _ = self.net_g_ms.eval()
45
+ utils.load_checkpoint(model, self.net_g_ms)
46
+
47
+ def get_text(self, text, hps, cleaned=False):
48
+ if cleaned:
49
+ text_norm = text_to_sequence(text, hps.symbols, [])
50
+ else:
51
+ text_norm = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
52
+ if hps.data.add_blank:
53
+ text_norm = commons.intersperse(text_norm, 0)
54
+ text_norm = LongTensor(text_norm)
55
+ return text_norm
56
+
57
+ def get_label_value(self, text, label, default, warning_name='value'):
58
+ value = re.search(rf'\[{label}=(.+?)\]', text)
59
+ if value:
60
+ try:
61
+ text = re.sub(rf'\[{label}=(.+?)\]', '', text, 1)
62
+ value = float(value.group(1))
63
+ except:
64
+ print(f'Invalid {warning_name}!')
65
+ sys.exit(1)
66
+ else:
67
+ value = default
68
+ return value, text
69
+
70
+ def ex_return(self, text, escape=False):
71
+ if escape:
72
+ return text.encode('unicode_escape').decode()
73
+ else:
74
+ return text
75
+
76
+ def return_speakers(self, escape=False):
77
+ return self.speakers
78
+
79
+ def get_label(self, text, label):
80
+ if f'[{label}]' in text:
81
+ return True, text.replace(f'[{label}]', '')
82
+ else:
83
+ return False, text
84
+
85
+ def generate(self, text=None, speaker_id=None, format=None, speed=1, audio_path=None, target_id=None, escape=False,
86
+ option=None, w2v2_folder=None):
87
+ if self.n_symbols != 0:
88
+ if not self.emotion_embedding:
89
+ length_scale, text = self.get_label_value(text, 'LENGTH', speed, 'length scale')
90
+ noise_scale, text = self.get_label_value(text, 'NOISE', 0.667, 'noise scale')
91
+ noise_scale_w, text = self.get_label_value(text, 'NOISEW', 0.8, 'deviation of noise')
92
+ cleaned, text = self.get_label(text, 'CLEANED')
93
+
94
+ stn_tst = self.get_text(text, self.hps_ms, cleaned=cleaned)
95
+ with no_grad():
96
+ x_tst = stn_tst.unsqueeze(0)
97
+ x_tst_lengths = LongTensor([stn_tst.size(0)])
98
+ sid = LongTensor([speaker_id])
99
+ audio = self.net_g_ms.infer(x_tst, x_tst_lengths, sid=sid,
100
+ noise_scale=noise_scale,
101
+ noise_scale_w=noise_scale_w,
102
+ length_scale=length_scale)[0][0, 0].data.cpu().float().numpy()
103
+
104
+ # else:
105
+ # w2v2_model = audonnx.load(os.path.dirname(w2v2_folder))
106
+ #
107
+ # if option == 'clean':
108
+ # self.ex_print(_clean_text(
109
+ # text, self.hps_ms.data.text_cleaners), escape)
110
+ #
111
+ # length_scale, text = self.get_label_value(
112
+ # text, 'LENGTH', 1, 'length scale')
113
+ # noise_scale, text = self.get_label_value(
114
+ # text, 'NOISE', 0.667, 'noise scale')
115
+ # noise_scale_w, text = self.get_label_value(
116
+ # text, 'NOISEW', 0.8, 'deviation of noise')
117
+ # cleaned, text = self.get_label(text, 'CLEANED')
118
+ #
119
+ # stn_tst = self.get_text(text, self.hps_ms, cleaned=cleaned)
120
+ #
121
+ # emotion_reference = input('Path of an emotion reference: ')
122
+ # if emotion_reference.endswith('.npy'):
123
+ # emotion = np.load(emotion_reference)
124
+ # emotion = FloatTensor(emotion).unsqueeze(0)
125
+ # else:
126
+ # audio16000, sampling_rate = librosa.load(
127
+ # emotion_reference, sr=16000, mono=True)
128
+ # emotion = w2v2_model(audio16000, sampling_rate)[
129
+ # 'hidden_states']
130
+ # emotion_reference = re.sub(
131
+ # r'\..*$', '', emotion_reference)
132
+ # np.save(emotion_reference, emotion.squeeze(0))
133
+ # emotion = FloatTensor(emotion)
134
+ #
135
+ #
136
+ # with no_grad():
137
+ # x_tst = stn_tst.unsqueeze(0)
138
+ # x_tst_lengths = LongTensor([stn_tst.size(0)])
139
+ # sid = LongTensor([speaker_id])
140
+ # audio = self.net_g_ms.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=noise_scale,
141
+ # noise_scale_w=noise_scale_w,
142
+ # length_scale=length_scale, emotion_embedding=emotion)[0][
143
+ # 0, 0].data.cpu().float().numpy()
144
+
145
+ # else:
146
+ # model = input('Path of a hubert-soft Model: ')
147
+ # from hubert_model import hubert_soft
148
+ # hubert = hubert_soft(model)
149
+
150
+ # if audio_path != '[VC]':
151
+ # if self.use_f0:
152
+ # audio, sampling_rate = librosa.load(
153
+ # audio_path, sr=self.hps_ms.data.sampling_rate, mono=True)
154
+ # audio16000 = librosa.resample(
155
+ # audio, orig_sr=sampling_rate, target_sr=16000)
156
+ # else:
157
+ # audio16000, sampling_rate = librosa.load(
158
+ # audio_path, sr=16000, mono=True)
159
+ #
160
+ # out_path = "H:/git/MoeGoe-Simple-API/upload/hubert.wav"
161
+ # length_scale, out_path = self.get_label_value(
162
+ # out_path, 'LENGTH', 1, 'length scale')
163
+ # noise_scale, out_path = self.get_label_value(
164
+ # out_path, 'NOISE', 0.1, 'noise scale')
165
+ # noise_scale_w, out_path = self.get_label_value(
166
+ # out_path, 'NOISEW', 0.1, 'deviation of noise')
167
+ #
168
+ # with inference_mode():
169
+ # units = hubert.units(FloatTensor(audio16000).unsqueeze(
170
+ # 0).unsqueeze(0)).squeeze(0).numpy()
171
+ # if self.use_f0:
172
+ # f0_scale, out_path = self.get_label_value(
173
+ # out_path, 'F0', 1, 'f0 scale')
174
+ # f0 = librosa.pyin(audio, sr=sampling_rate,
175
+ # fmin=librosa.note_to_hz('C0'),
176
+ # fmax=librosa.note_to_hz('C7'),
177
+ # frame_length=1780)[0]
178
+ # target_length = len(units[:, 0])
179
+ # f0 = np.nan_to_num(np.interp(np.arange(0, len(f0) * target_length, len(f0)) / target_length,
180
+ # np.arange(0, len(f0)), f0)) * f0_scale
181
+ # units[:, 0] = f0 / 10
182
+ #
183
+ # stn_tst = FloatTensor(units)
184
+ # with no_grad():
185
+ # x_tst = stn_tst.unsqueeze(0)
186
+ # x_tst_lengths = LongTensor([stn_tst.size(0)])
187
+ # sid = LongTensor([target_id])
188
+ # audio = self.net_g_ms.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=noise_scale,
189
+ # noise_scale_w=noise_scale_w, length_scale=length_scale)[0][
190
+ # 0, 0].data.float().numpy()
191
+
192
+ with BytesIO() as f:
193
+ fname = str(uuid.uuid1())
194
+
195
+ if format == 'ogg':
196
+ write(f, self.hps_ms.data.sampling_rate, audio)
197
+ with BytesIO() as o:
198
+ utils.wav2ogg(f, o)
199
+ return BytesIO(o.getvalue()), "audio/ogg", fname + ".ogg"
200
+ elif format == 'silk':
201
+ file_path = self.out_path + "/" + fname + ".wav"
202
+ write(file_path, 24000, audio)
203
+ silk_path = utils.convert_to_silk(file_path)
204
+ os.remove(file_path)
205
+ return silk_path, "audio/silk", fname + ".silk"
206
+ else:
207
+ write(f, self.hps_ms.data.sampling_rate, audio)
208
+ return BytesIO(f.getvalue()), "audio/wav", fname + ".wav"
209
+
210
+ def voice_conversion(self, audio_path, original_id, target_id):
211
+
212
+ audio = utils.load_audio_to_torch(
213
+ audio_path, self.hps_ms.data.sampling_rate)
214
+
215
+ y = audio.unsqueeze(0)
216
+
217
+ spec = spectrogram_torch(y, self.hps_ms.data.filter_length,
218
+ self.hps_ms.data.sampling_rate, self.hps_ms.data.hop_length,
219
+ self.hps_ms.data.win_length,
220
+ center=False)
221
+ spec_lengths = LongTensor([spec.size(-1)])
222
+ sid_src = LongTensor([original_id])
223
+
224
+ with no_grad():
225
+ sid_tgt = LongTensor([target_id])
226
+ audio = self.net_g_ms.voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt)[
227
+ 0][0, 0].data.cpu().float().numpy()
228
+
229
+ with BytesIO() as f:
230
+ write(f, self.hps_ms.data.sampling_rate, audio)
231
+ return BytesIO(f.getvalue())