Spaces:
Runtime error
Runtime error
XiaoHei Studio
commited on
Commit
•
8907ed4
1
Parent(s):
abc4e5e
Upload 13 files
Browse files- filelists/test.txt +4 -0
- filelists/train.txt +566 -0
- filelists/val.txt +2 -0
- inference/__init__.py +0 -0
- inference/__pycache__/__init__.cpython-38.pyc +0 -0
- inference/__pycache__/infer_tool.cpython-38.pyc +0 -0
- inference/__pycache__/infer_tool_webui.cpython-38.pyc +0 -0
- inference/__pycache__/slicer.cpython-38.pyc +0 -0
- inference/chunks_temp.json +1 -0
- inference/infer_tool.py +546 -0
- inference/infer_tool_grad.py +156 -0
- inference/infer_tool_webui.py +547 -0
- inference/slicer.py +142 -0
filelists/test.txt
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
./dataset/44k/taffy/000562.wav
|
| 2 |
+
./dataset/44k/nyaru/000011.wav
|
| 3 |
+
./dataset/44k/nyaru/000008.wav
|
| 4 |
+
./dataset/44k/taffy/000563.wav
|
filelists/train.txt
ADDED
|
@@ -0,0 +1,566 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
./dataset/44k/chino_v7/chino_diff_aug_251.wav
|
| 2 |
+
./dataset/44k/chino_v7/chino_reprocess_744.wav
|
| 3 |
+
./dataset/44k/chino_v7/chino_diff_aug_2.wav
|
| 4 |
+
./dataset/44k/chino_v7/chino_reprocess_872.wav
|
| 5 |
+
./dataset/44k/chino_v7/chino_diff_aug_382_0.wav
|
| 6 |
+
./dataset/44k/chino_v7/chino_diff_aug_262.wav
|
| 7 |
+
./dataset/44k/chino_v7/chino_diff_aug_264.wav
|
| 8 |
+
./dataset/44k/chino_v7/chino_diff_aug_268.wav
|
| 9 |
+
./dataset/44k/chino_v7/chino_diff_aug_64.wav
|
| 10 |
+
./dataset/44k/chino_v7/chino_diff_aug_323.wav
|
| 11 |
+
./dataset/44k/chino_v7/chino_diff_aug_316.wav
|
| 12 |
+
./dataset/44k/chino_v7/chino_reprocess_801.wav
|
| 13 |
+
./dataset/44k/chino_v7/chino_diff_aug_29.wav
|
| 14 |
+
./dataset/44k/chino_v7/chino_diff_aug_87_3.wav
|
| 15 |
+
./dataset/44k/chino_v7/chino_reprocess_766.wav
|
| 16 |
+
./dataset/44k/chino_v7/chino_diff_aug_91.wav
|
| 17 |
+
./dataset/44k/chino_v7/chino_diff_aug_217_2.wav
|
| 18 |
+
./dataset/44k/chino_v7/chino_diff_aug_88.wav
|
| 19 |
+
./dataset/44k/chino_v7/chino_diff_aug_61.wav
|
| 20 |
+
./dataset/44k/chino_v7/chino_reprocess_828.wav
|
| 21 |
+
./dataset/44k/chino_v7/chino_diff_aug_109.wav
|
| 22 |
+
./dataset/44k/chino_v7/chino_diff_aug_306.wav
|
| 23 |
+
./dataset/44k/chino_v7/chino_diff_aug_40.wav
|
| 24 |
+
./dataset/44k/chino_v7/chino_diff_aug_220_2.wav
|
| 25 |
+
./dataset/44k/chino_v7/chino_diff_aug_310_2.wav
|
| 26 |
+
./dataset/44k/chino_v7/chino_diff_aug_3.wav
|
| 27 |
+
./dataset/44k/chino_v7/chino_diff_aug_293.wav
|
| 28 |
+
./dataset/44k/chino_v7/chino_diff_aug_176.wav
|
| 29 |
+
./dataset/44k/chino_v7/chino_diff_aug_93.wav
|
| 30 |
+
./dataset/44k/chino_v7/chino_diff_aug_140_4.wav
|
| 31 |
+
./dataset/44k/chino_v7/chino_diff_aug_174_0.wav
|
| 32 |
+
./dataset/44k/chino_v7/chino_diff_aug_212.wav
|
| 33 |
+
./dataset/44k/chino_v7/chino_diff_aug_296.wav
|
| 34 |
+
./dataset/44k/chino_v7/chino_diff_aug_300.wav
|
| 35 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_6.wav
|
| 36 |
+
./dataset/44k/chino_v7/chino_reprocess_819.wav
|
| 37 |
+
./dataset/44k/chino_v7/chino_diff_aug_382_1.wav
|
| 38 |
+
./dataset/44k/chino_v7/chino_diff_aug_225.wav
|
| 39 |
+
./dataset/44k/chino_v7/chino_diff_aug_139_1.wav
|
| 40 |
+
./dataset/44k/chino_v7/chino_diff_aug_149_2.wav
|
| 41 |
+
./dataset/44k/chino_v7/chino_diff_aug_320.wav
|
| 42 |
+
./dataset/44k/chino_v7/chino_reprocess_844.wav
|
| 43 |
+
./dataset/44k/chino_v7/chino_diff_aug_398.wav
|
| 44 |
+
./dataset/44k/chino_v7/chino_reprocess_747.wav
|
| 45 |
+
./dataset/44k/chino_v7/chino_diff_aug_213_1.wav
|
| 46 |
+
./dataset/44k/chino_v7/chino_diff_aug_194.wav
|
| 47 |
+
./dataset/44k/chino_v7/chino_diff_aug_167_0.wav
|
| 48 |
+
./dataset/44k/chino_v7/chino_diff_aug_393_0.wav
|
| 49 |
+
./dataset/44k/chino_v7/chino_diff_aug_62.wav
|
| 50 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_4.wav
|
| 51 |
+
./dataset/44k/chino_v7/chino_reprocess_875.wav
|
| 52 |
+
./dataset/44k/chino_v7/chino_diff_aug_111.wav
|
| 53 |
+
./dataset/44k/chino_v7/chino_reprocess_817.wav
|
| 54 |
+
./dataset/44k/chino_v7/chino_diff_aug_39.wav
|
| 55 |
+
./dataset/44k/chino_v7/chino_diff_aug_353_3.wav
|
| 56 |
+
./dataset/44k/chino_v7/chino_diff_aug_329.wav
|
| 57 |
+
./dataset/44k/chino_v7/chino_diff_aug_265.wav
|
| 58 |
+
./dataset/44k/chino_v7/chino_reprocess_831.wav
|
| 59 |
+
./dataset/44k/chino_v7/chino_diff_aug_42_1.wav
|
| 60 |
+
./dataset/44k/chino_v7/chino_diff_aug_330.wav
|
| 61 |
+
./dataset/44k/chino_v7/chino_diff_aug_207.wav
|
| 62 |
+
./dataset/44k/chino_v7/chino_diff_aug_334.wav
|
| 63 |
+
./dataset/44k/chino_v7/chino_diff_aug_6_0.wav
|
| 64 |
+
./dataset/44k/chino_v7/chino_diff_aug_18.wav
|
| 65 |
+
./dataset/44k/chino_v7/chino_diff_aug_342.wav
|
| 66 |
+
./dataset/44k/chino_v7/chino_diff_aug_397.wav
|
| 67 |
+
./dataset/44k/chino_v7/chino_diff_aug_164.wav
|
| 68 |
+
./dataset/44k/chino_v7/chino_diff_aug_313.wav
|
| 69 |
+
./dataset/44k/chino_v7/chino_diff_aug_211.wav
|
| 70 |
+
./dataset/44k/chino_v7/chino_reprocess_836.wav
|
| 71 |
+
./dataset/44k/chino_v7/chino_diff_aug_339.wav
|
| 72 |
+
./dataset/44k/chino_v7/chino_diff_aug_157_3.wav
|
| 73 |
+
./dataset/44k/chino_v7/chino_reprocess_853.wav
|
| 74 |
+
./dataset/44k/chino_v7/chino_reprocess_803.wav
|
| 75 |
+
./dataset/44k/chino_v7/chino_diff_aug_189.wav
|
| 76 |
+
./dataset/44k/chino_v7/chino_diff_aug_248.wav
|
| 77 |
+
./dataset/44k/chino_v7/chino_diff_aug_277.wav
|
| 78 |
+
./dataset/44k/chino_v7/chino_reprocess_772.wav
|
| 79 |
+
./dataset/44k/chino_v7/chino_reprocess_778.wav
|
| 80 |
+
./dataset/44k/chino_v7/chino_diff_aug_285.wav
|
| 81 |
+
./dataset/44k/chino_v7/chino_diff_aug_301.wav
|
| 82 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_3.wav
|
| 83 |
+
./dataset/44k/chino_v7/chino_diff_aug_228.wav
|
| 84 |
+
./dataset/44k/chino_v7/chino_diff_aug_68.wav
|
| 85 |
+
./dataset/44k/chino_v7/chino_reprocess_770.wav
|
| 86 |
+
./dataset/44k/chino_v7/chino_reprocess_730.wav
|
| 87 |
+
./dataset/44k/chino_v7/chino_diff_aug_324.wav
|
| 88 |
+
./dataset/44k/chino_v7/chino_diff_aug_340.wav
|
| 89 |
+
./dataset/44k/chino_v7/chino_diff_aug_45.wav
|
| 90 |
+
./dataset/44k/chino_v7/chino_reprocess_716.wav
|
| 91 |
+
./dataset/44k/chino_v7/chino_diff_aug_19.wav
|
| 92 |
+
./dataset/44k/chino_v7/chino_reprocess_724.wav
|
| 93 |
+
./dataset/44k/chino_v7/chino_reprocess_882.wav
|
| 94 |
+
./dataset/44k/chino_v7/chino_reprocess_809.wav
|
| 95 |
+
./dataset/44k/chino_v7/chino_diff_aug_354.wav
|
| 96 |
+
./dataset/44k/chino_v7/chino_diff_aug_49.wav
|
| 97 |
+
./dataset/44k/chino_v7/chino_diff_aug_242.wav
|
| 98 |
+
./dataset/44k/chino_v7/chino_reprocess_810.wav
|
| 99 |
+
./dataset/44k/chino_v7/chino_diff_aug_96_0.wav
|
| 100 |
+
./dataset/44k/chino_v7/chino_reprocess_798.wav
|
| 101 |
+
./dataset/44k/chino_v7/chino_diff_aug_14.wav
|
| 102 |
+
./dataset/44k/chino_v7/chino_diff_aug_372.wav
|
| 103 |
+
./dataset/44k/chino_v7/chino_diff_aug_84.wav
|
| 104 |
+
./dataset/44k/chino_v7/chino_diff_aug_302.wav
|
| 105 |
+
./dataset/44k/chino_v7/chino_diff_aug_256.wav
|
| 106 |
+
./dataset/44k/chino_v7/chino_reprocess_751.wav
|
| 107 |
+
./dataset/44k/chino_v7/chino_diff_aug_104.wav
|
| 108 |
+
./dataset/44k/chino_v7/chino_diff_aug_125_1.wav
|
| 109 |
+
./dataset/44k/chino_v7/chino_diff_aug_315_0.wav
|
| 110 |
+
./dataset/44k/chino_v7/chino_diff_aug_186_0.wav
|
| 111 |
+
./dataset/44k/chino_v7/chino_diff_aug_275.wav
|
| 112 |
+
./dataset/44k/chino_v7/chino_diff_aug_59.wav
|
| 113 |
+
./dataset/44k/chino_v7/chino_diff_aug_382_2.wav
|
| 114 |
+
./dataset/44k/chino_v7/chino_diff_aug_297.wav
|
| 115 |
+
./dataset/44k/chino_v7/chino_diff_aug_254.wav
|
| 116 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_2.wav
|
| 117 |
+
./dataset/44k/chino_v7/chino_diff_aug_233.wav
|
| 118 |
+
./dataset/44k/chino_v7/chino_diff_aug_258.wav
|
| 119 |
+
./dataset/44k/chino_v7/chino_diff_aug_360_0.wav
|
| 120 |
+
./dataset/44k/chino_v7/chino_reprocess_779.wav
|
| 121 |
+
./dataset/44k/chino_v7/chino_diff_aug_170_0.wav
|
| 122 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_1.wav
|
| 123 |
+
./dataset/44k/chino_v7/chino_diff_aug_195.wav
|
| 124 |
+
./dataset/44k/chino_v7/chino_diff_aug_244.wav
|
| 125 |
+
./dataset/44k/chino_v7/chino_diff_aug_353_1.wav
|
| 126 |
+
./dataset/44k/chino_v7/chino_reprocess_799.wav
|
| 127 |
+
./dataset/44k/chino_v7/chino_diff_aug_393_2.wav
|
| 128 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_5.wav
|
| 129 |
+
./dataset/44k/chino_v7/chino_diff_aug_74.wav
|
| 130 |
+
./dataset/44k/chino_v7/chino_diff_aug_373.wav
|
| 131 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_2.wav
|
| 132 |
+
./dataset/44k/chino_v7/chino_diff_aug_304.wav
|
| 133 |
+
./dataset/44k/chino_v7/chino_diff_aug_114.wav
|
| 134 |
+
./dataset/44k/chino_v7/chino_diff_aug_31.wav
|
| 135 |
+
./dataset/44k/chino_v7/chino_diff_aug_217_1.wav
|
| 136 |
+
./dataset/44k/chino_v7/chino_diff_aug_200.wav
|
| 137 |
+
./dataset/44k/chino_v7/chino_diff_aug_367_2.wav
|
| 138 |
+
./dataset/44k/chino_v7/chino_diff_aug_157_1.wav
|
| 139 |
+
./dataset/44k/chino_v7/chino_reprocess_852.wav
|
| 140 |
+
./dataset/44k/chino_v7/chino_diff_aug_379.wav
|
| 141 |
+
./dataset/44k/chino_v7/chino_reprocess_726.wav
|
| 142 |
+
./dataset/44k/chino_v7/chino_diff_aug_37.wav
|
| 143 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_0.wav
|
| 144 |
+
./dataset/44k/chino_v7/chino_diff_aug_198.wav
|
| 145 |
+
./dataset/44k/chino_v7/chino_reprocess_859.wav
|
| 146 |
+
./dataset/44k/chino_v7/chino_reprocess_841.wav
|
| 147 |
+
./dataset/44k/chino_v7/chino_reprocess_742.wav
|
| 148 |
+
./dataset/44k/chino_v7/chino_diff_aug_43.wav
|
| 149 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_4.wav
|
| 150 |
+
./dataset/44k/chino_v7/chino_reprocess_717.wav
|
| 151 |
+
./dataset/44k/chino_v7/chino_diff_aug_222.wav
|
| 152 |
+
./dataset/44k/chino_v7/chino_reprocess_722.wav
|
| 153 |
+
./dataset/44k/chino_v7/chino_diff_aug_294.wav
|
| 154 |
+
./dataset/44k/chino_v7/chino_diff_aug_351.wav
|
| 155 |
+
./dataset/44k/chino_v7/chino_diff_aug_333.wav
|
| 156 |
+
./dataset/44k/chino_v7/chino_diff_aug_137_0.wav
|
| 157 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_4.wav
|
| 158 |
+
./dataset/44k/chino_v7/chino_diff_aug_184.wav
|
| 159 |
+
./dataset/44k/chino_v7/chino_diff_aug_328.wav
|
| 160 |
+
./dataset/44k/chino_v7/chino_reprocess_874.wav
|
| 161 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_0.wav
|
| 162 |
+
./dataset/44k/chino_v7/chino_diff_aug_170_3.wav
|
| 163 |
+
./dataset/44k/chino_v7/chino_diff_aug_249.wav
|
| 164 |
+
./dataset/44k/chino_v7/chino_reprocess_752.wav
|
| 165 |
+
./dataset/44k/chino_v7/chino_diff_aug_75.wav
|
| 166 |
+
./dataset/44k/chino_v7/chino_diff_aug_353_0.wav
|
| 167 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_2.wav
|
| 168 |
+
./dataset/44k/chino_v7/chino_diff_aug_377.wav
|
| 169 |
+
./dataset/44k/chino_v7/chino_diff_aug_106.wav
|
| 170 |
+
./dataset/44k/chino_v7/chino_diff_aug_229_2.wav
|
| 171 |
+
./dataset/44k/chino_v7/chino_diff_aug_388.wav
|
| 172 |
+
./dataset/44k/chino_v7/chino_diff_aug_234_0.wav
|
| 173 |
+
./dataset/44k/chino_v7/chino_diff_aug_125_0.wav
|
| 174 |
+
./dataset/44k/chino_v7/chino_diff_aug_282.wav
|
| 175 |
+
./dataset/44k/chino_v7/chino_diff_aug_292.wav
|
| 176 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_1.wav
|
| 177 |
+
./dataset/44k/chino_v7/chino_diff_aug_42_0.wav
|
| 178 |
+
./dataset/44k/chino_v7/chino_diff_aug_202_2.wav
|
| 179 |
+
./dataset/44k/chino_v7/chino_reprocess_862.wav
|
| 180 |
+
./dataset/44k/chino_v7/chino_diff_aug_308.wav
|
| 181 |
+
./dataset/44k/chino_v7/chino_diff_aug_400_0.wav
|
| 182 |
+
./dataset/44k/chino_v7/chino_diff_aug_41.wav
|
| 183 |
+
./dataset/44k/chino_v7/chino_diff_aug_325.wav
|
| 184 |
+
./dataset/44k/chino_v7/chino_reprocess_788.wav
|
| 185 |
+
./dataset/44k/chino_v7/chino_diff_aug_310_0.wav
|
| 186 |
+
./dataset/44k/chino_v7/chino_diff_aug_226.wav
|
| 187 |
+
./dataset/44k/chino_v7/chino_reprocess_795.wav
|
| 188 |
+
./dataset/44k/chino_v7/chino_diff_aug_270.wav
|
| 189 |
+
./dataset/44k/chino_v7/chino_diff_aug_153.wav
|
| 190 |
+
./dataset/44k/chino_v7/chino_diff_aug_403_1.wav
|
| 191 |
+
./dataset/44k/chino_v7/chino_reprocess_858.wav
|
| 192 |
+
./dataset/44k/chino_v7/chino_diff_aug_12_0.wav
|
| 193 |
+
./dataset/44k/chino_v7/chino_diff_aug_403_3.wav
|
| 194 |
+
./dataset/44k/chino_v7/chino_diff_aug_130.wav
|
| 195 |
+
./dataset/44k/chino_v7/chino_diff_aug_63.wav
|
| 196 |
+
./dataset/44k/chino_v7/chino_diff_aug_108.wav
|
| 197 |
+
./dataset/44k/chino_v7/chino_diff_aug_134.wav
|
| 198 |
+
./dataset/44k/chino_v7/chino_diff_aug_276.wav
|
| 199 |
+
./dataset/44k/chino_v7/chino_diff_aug_183.wav
|
| 200 |
+
./dataset/44k/chino_v7/chino_diff_aug_100_0.wav
|
| 201 |
+
./dataset/44k/chino_v7/chino_diff_aug_159.wav
|
| 202 |
+
./dataset/44k/chino_v7/chino_diff_aug_400_2.wav
|
| 203 |
+
./dataset/44k/chino_v7/chino_diff_aug_386.wav
|
| 204 |
+
./dataset/44k/chino_v7/chino_reprocess_781.wav
|
| 205 |
+
./dataset/44k/chino_v7/chino_reprocess_750.wav
|
| 206 |
+
./dataset/44k/chino_v7/chino_diff_aug_229_1.wav
|
| 207 |
+
./dataset/44k/chino_v7/chino_diff_aug_7.wav
|
| 208 |
+
./dataset/44k/chino_v7/chino_reprocess_805.wav
|
| 209 |
+
./dataset/44k/chino_v7/chino_diff_aug_384.wav
|
| 210 |
+
./dataset/44k/chino_v7/chino_diff_aug_171.wav
|
| 211 |
+
./dataset/44k/chino_v7/chino_diff_aug_220_0.wav
|
| 212 |
+
./dataset/44k/chino_v7/chino_reprocess_870.wav
|
| 213 |
+
./dataset/44k/chino_v7/chino_diff_aug_165.wav
|
| 214 |
+
./dataset/44k/chino_v7/chino_diff_aug_115.wav
|
| 215 |
+
./dataset/44k/chino_v7/chino_diff_aug_191.wav
|
| 216 |
+
./dataset/44k/chino_v7/chino_reprocess_849.wav
|
| 217 |
+
./dataset/44k/chino_v7/chino_diff_aug_185.wav
|
| 218 |
+
./dataset/44k/chino_v7/chino_diff_aug_400_1.wav
|
| 219 |
+
./dataset/44k/chino_v7/chino_diff_aug_341.wav
|
| 220 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_3.wav
|
| 221 |
+
./dataset/44k/chino_v7/chino_diff_aug_234_2.wav
|
| 222 |
+
./dataset/44k/chino_v7/chino_diff_aug_112.wav
|
| 223 |
+
./dataset/44k/chino_v7/chino_diff_aug_252.wav
|
| 224 |
+
./dataset/44k/chino_v7/chino_diff_aug_396.wav
|
| 225 |
+
./dataset/44k/chino_v7/chino_reprocess_787.wav
|
| 226 |
+
./dataset/44k/chino_v7/chino_diff_aug_202_0.wav
|
| 227 |
+
./dataset/44k/chino_v7/chino_diff_aug_378.wav
|
| 228 |
+
./dataset/44k/chino_v7/chino_diff_aug_219.wav
|
| 229 |
+
./dataset/44k/chino_v7/chino_diff_aug_55.wav
|
| 230 |
+
./dataset/44k/chino_v7/chino_diff_aug_8_2.wav
|
| 231 |
+
./dataset/44k/chino_v7/chino_diff_aug_148_1.wav
|
| 232 |
+
./dataset/44k/chino_v7/chino_diff_aug_11.wav
|
| 233 |
+
./dataset/44k/chino_v7/chino_diff_aug_288.wav
|
| 234 |
+
./dataset/44k/chino_v7/chino_reprocess_823.wav
|
| 235 |
+
./dataset/44k/chino_v7/chino_diff_aug_170_2.wav
|
| 236 |
+
./dataset/44k/chino_v7/chino_diff_aug_283.wav
|
| 237 |
+
./dataset/44k/chino_v7/chino_reprocess_741.wav
|
| 238 |
+
./dataset/44k/chino_v7/chino_diff_aug_32.wav
|
| 239 |
+
./dataset/44k/chino_v7/chino_diff_aug_299.wav
|
| 240 |
+
./dataset/44k/chino_v7/chino_diff_aug_161.wav
|
| 241 |
+
./dataset/44k/chino_v7/chino_reprocess_736.wav
|
| 242 |
+
./dataset/44k/chino_v7/chino_diff_aug_73.wav
|
| 243 |
+
./dataset/44k/chino_v7/chino_diff_aug_326.wav
|
| 244 |
+
./dataset/44k/chino_v7/chino_diff_aug_12_1.wav
|
| 245 |
+
./dataset/44k/chino_v7/chino_reprocess_867.wav
|
| 246 |
+
./dataset/44k/chino_v7/chino_diff_aug_124.wav
|
| 247 |
+
./dataset/44k/chino_v7/chino_reprocess_791.wav
|
| 248 |
+
./dataset/44k/chino_v7/chino_diff_aug_128.wav
|
| 249 |
+
./dataset/44k/chino_v7/chino_diff_aug_214.wav
|
| 250 |
+
./dataset/44k/chino_v7/chino_diff_aug_403_2.wav
|
| 251 |
+
./dataset/44k/chino_v7/chino_reprocess_725.wav
|
| 252 |
+
./dataset/44k/chino_v7/chino_diff_aug_118.wav
|
| 253 |
+
./dataset/44k/chino_v7/chino_reprocess_790.wav
|
| 254 |
+
./dataset/44k/chino_v7/chino_reprocess_825.wav
|
| 255 |
+
./dataset/44k/chino_v7/chino_diff_aug_217_0.wav
|
| 256 |
+
./dataset/44k/chino_v7/chino_diff_aug_54.wav
|
| 257 |
+
./dataset/44k/chino_v7/chino_reprocess_780.wav
|
| 258 |
+
./dataset/44k/chino_v7/chino_diff_aug_12_3.wav
|
| 259 |
+
./dataset/44k/chino_v7/chino_diff_aug_266.wav
|
| 260 |
+
./dataset/44k/chino_v7/chino_reprocess_762.wav
|
| 261 |
+
./dataset/44k/chino_v7/chino_diff_aug_65.wav
|
| 262 |
+
./dataset/44k/chino_v7/chino_diff_aug_0.wav
|
| 263 |
+
./dataset/44k/chino_v7/chino_diff_aug_392.wav
|
| 264 |
+
./dataset/44k/chino_v7/chino_diff_aug_139_0.wav
|
| 265 |
+
./dataset/44k/chino_v7/chino_diff_aug_400_3.wav
|
| 266 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_3.wav
|
| 267 |
+
./dataset/44k/chino_v7/chino_reprocess_814.wav
|
| 268 |
+
./dataset/44k/chino_v7/chino_diff_aug_348.wav
|
| 269 |
+
./dataset/44k/chino_v7/chino_diff_aug_13.wav
|
| 270 |
+
./dataset/44k/chino_v7/chino_diff_aug_48.wav
|
| 271 |
+
./dataset/44k/chino_v7/chino_diff_aug_52.wav
|
| 272 |
+
./dataset/44k/chino_v7/chino_reprocess_784.wav
|
| 273 |
+
./dataset/44k/chino_v7/chino_diff_aug_314.wav
|
| 274 |
+
./dataset/44k/chino_v7/chino_reprocess_804.wav
|
| 275 |
+
./dataset/44k/chino_v7/chino_diff_aug_399.wav
|
| 276 |
+
./dataset/44k/chino_v7/chino_reprocess_827.wav
|
| 277 |
+
./dataset/44k/chino_v7/chino_diff_aug_369_3.wav
|
| 278 |
+
./dataset/44k/chino_v7/chino_diff_aug_281.wav
|
| 279 |
+
./dataset/44k/chino_v7/chino_reprocess_835.wav
|
| 280 |
+
./dataset/44k/chino_v7/chino_diff_aug_46.wav
|
| 281 |
+
./dataset/44k/chino_v7/chino_diff_aug_144.wav
|
| 282 |
+
./dataset/44k/chino_v7/chino_diff_aug_188.wav
|
| 283 |
+
./dataset/44k/chino_v7/chino_diff_aug_121.wav
|
| 284 |
+
./dataset/44k/chino_v7/chino_diff_aug_160.wav
|
| 285 |
+
./dataset/44k/chino_v7/chino_diff_aug_205.wav
|
| 286 |
+
./dataset/44k/chino_v7/chino_diff_aug_4.wav
|
| 287 |
+
./dataset/44k/chino_v7/chino_diff_aug_103.wav
|
| 288 |
+
./dataset/44k/chino_v7/chino_diff_aug_71.wav
|
| 289 |
+
./dataset/44k/chino_v7/chino_reprocess_785.wav
|
| 290 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_2.wav
|
| 291 |
+
./dataset/44k/chino_v7/chino_diff_aug_261.wav
|
| 292 |
+
./dataset/44k/chino_v7/chino_reprocess_829.wav
|
| 293 |
+
./dataset/44k/chino_v7/chino_diff_aug_289.wav
|
| 294 |
+
./dataset/44k/chino_v7/chino_diff_aug_208.wav
|
| 295 |
+
./dataset/44k/chino_v7/chino_diff_aug_332.wav
|
| 296 |
+
./dataset/44k/chino_v7/chino_diff_aug_162.wav
|
| 297 |
+
./dataset/44k/chino_v7/chino_diff_aug_143_0.wav
|
| 298 |
+
./dataset/44k/chino_v7/chino_diff_aug_201.wav
|
| 299 |
+
./dataset/44k/chino_v7/chino_reprocess_885.wav
|
| 300 |
+
./dataset/44k/chino_v7/chino_diff_aug_369_0.wav
|
| 301 |
+
./dataset/44k/chino_v7/chino_reprocess_794.wav
|
| 302 |
+
./dataset/44k/chino_v7/chino_diff_aug_202_3.wav
|
| 303 |
+
./dataset/44k/chino_v7/chino_reprocess_857.wav
|
| 304 |
+
./dataset/44k/chino_v7/chino_diff_aug_257.wav
|
| 305 |
+
./dataset/44k/chino_v7/chino_diff_aug_163_1.wav
|
| 306 |
+
./dataset/44k/chino_v7/chino_diff_aug_267.wav
|
| 307 |
+
./dataset/44k/chino_v7/chino_diff_aug_36_1.wav
|
| 308 |
+
./dataset/44k/chino_v7/chino_diff_aug_387.wav
|
| 309 |
+
./dataset/44k/chino_v7/chino_diff_aug_327.wav
|
| 310 |
+
./dataset/44k/chino_v7/chino_reprocess_806.wav
|
| 311 |
+
./dataset/44k/chino_v7/chino_reprocess_822.wav
|
| 312 |
+
./dataset/44k/chino_v7/chino_diff_aug_238_2.wav
|
| 313 |
+
./dataset/44k/chino_v7/chino_diff_aug_168.wav
|
| 314 |
+
./dataset/44k/chino_v7/chino_reprocess_721.wav
|
| 315 |
+
./dataset/44k/chino_v7/chino_diff_aug_138.wav
|
| 316 |
+
./dataset/44k/chino_v7/chino_diff_aug_15.wav
|
| 317 |
+
./dataset/44k/chino_v7/chino_diff_aug_352.wav
|
| 318 |
+
./dataset/44k/chino_v7/chino_reprocess_854.wav
|
| 319 |
+
./dataset/44k/chino_v7/chino_diff_aug_336.wav
|
| 320 |
+
./dataset/44k/chino_v7/chino_diff_aug_38.wav
|
| 321 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_0.wav
|
| 322 |
+
./dataset/44k/chino_v7/chino_diff_aug_53.wav
|
| 323 |
+
./dataset/44k/chino_v7/chino_reprocess_887.wav
|
| 324 |
+
./dataset/44k/chino_v7/chino_reprocess_846.wav
|
| 325 |
+
./dataset/44k/chino_v7/chino_reprocess_776.wav
|
| 326 |
+
./dataset/44k/chino_v7/chino_reprocess_848.wav
|
| 327 |
+
./dataset/44k/chino_v7/chino_diff_aug_303.wav
|
| 328 |
+
./dataset/44k/chino_v7/chino_diff_aug_364.wav
|
| 329 |
+
./dataset/44k/chino_v7/chino_reprocess_843.wav
|
| 330 |
+
./dataset/44k/chino_v7/chino_diff_aug_231_2.wav
|
| 331 |
+
./dataset/44k/chino_v7/chino_diff_aug_393_1.wav
|
| 332 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_1.wav
|
| 333 |
+
./dataset/44k/chino_v7/chino_diff_aug_382_3.wav
|
| 334 |
+
./dataset/44k/chino_v7/chino_diff_aug_147.wav
|
| 335 |
+
./dataset/44k/chino_v7/chino_diff_aug_140_0.wav
|
| 336 |
+
./dataset/44k/chino_v7/chino_diff_aug_243.wav
|
| 337 |
+
./dataset/44k/chino_v7/chino_reprocess_889.wav
|
| 338 |
+
./dataset/44k/chino_v7/chino_reprocess_877.wav
|
| 339 |
+
./dataset/44k/chino_v7/chino_diff_aug_158.wav
|
| 340 |
+
./dataset/44k/chino_v7/chino_diff_aug_356.wav
|
| 341 |
+
./dataset/44k/chino_v7/chino_diff_aug_286.wav
|
| 342 |
+
./dataset/44k/chino_v7/chino_diff_aug_10.wav
|
| 343 |
+
./dataset/44k/chino_v7/chino_diff_aug_360_2.wav
|
| 344 |
+
./dataset/44k/chino_v7/chino_reprocess_796.wav
|
| 345 |
+
./dataset/44k/chino_v7/chino_diff_aug_23.wav
|
| 346 |
+
./dataset/44k/chino_v7/chino_reprocess_861.wav
|
| 347 |
+
./dataset/44k/chino_v7/chino_reprocess_869.wav
|
| 348 |
+
./dataset/44k/chino_v7/chino_diff_aug_202_1.wav
|
| 349 |
+
./dataset/44k/chino_v7/chino_diff_aug_319.wav
|
| 350 |
+
./dataset/44k/chino_v7/chino_reprocess_864.wav
|
| 351 |
+
./dataset/44k/chino_v7/chino_reprocess_826.wav
|
| 352 |
+
./dataset/44k/chino_v7/chino_diff_aug_175.wav
|
| 353 |
+
./dataset/44k/chino_v7/chino_diff_aug_9.wav
|
| 354 |
+
./dataset/44k/chino_v7/chino_diff_aug_36_0.wav
|
| 355 |
+
./dataset/44k/chino_v7/chino_diff_aug_206.wav
|
| 356 |
+
./dataset/44k/chino_v7/chino_reprocess_833.wav
|
| 357 |
+
./dataset/44k/chino_v7/chino_diff_aug_355.wav
|
| 358 |
+
./dataset/44k/chino_v7/chino_diff_aug_120.wav
|
| 359 |
+
./dataset/44k/chino_v7/chino_reprocess_881.wav
|
| 360 |
+
./dataset/44k/chino_v7/chino_diff_aug_215.wav
|
| 361 |
+
./dataset/44k/chino_v7/chino_diff_aug_234_1.wav
|
| 362 |
+
./dataset/44k/chino_v7/chino_diff_aug_129.wav
|
| 363 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_6.wav
|
| 364 |
+
./dataset/44k/chino_v7/chino_reprocess_868.wav
|
| 365 |
+
./dataset/44k/chino_v7/chino_diff_aug_167_1.wav
|
| 366 |
+
./dataset/44k/chino_v7/chino_diff_aug_156.wav
|
| 367 |
+
./dataset/44k/chino_v7/chino_diff_aug_298.wav
|
| 368 |
+
./dataset/44k/chino_v7/chino_diff_aug_51.wav
|
| 369 |
+
./dataset/44k/chino_v7/chino_reprocess_775.wav
|
| 370 |
+
./dataset/44k/chino_v7/chino_reprocess_839.wav
|
| 371 |
+
./dataset/44k/chino_v7/chino_diff_aug_117.wav
|
| 372 |
+
./dataset/44k/chino_v7/chino_diff_aug_247.wav
|
| 373 |
+
./dataset/44k/chino_v7/chino_diff_aug_123.wav
|
| 374 |
+
./dataset/44k/chino_v7/chino_reprocess_731.wav
|
| 375 |
+
./dataset/44k/chino_v7/chino_diff_aug_231_1.wav
|
| 376 |
+
./dataset/44k/chino_v7/chino_diff_aug_170_4.wav
|
| 377 |
+
./dataset/44k/chino_v7/chino_diff_aug_6_2.wav
|
| 378 |
+
./dataset/44k/chino_v7/chino_diff_aug_237_2.wav
|
| 379 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_7.wav
|
| 380 |
+
./dataset/44k/chino_v7/chino_diff_aug_146.wav
|
| 381 |
+
./dataset/44k/chino_v7/chino_diff_aug_271.wav
|
| 382 |
+
./dataset/44k/chino_v7/chino_diff_aug_403_0.wav
|
| 383 |
+
./dataset/44k/chino_v7/chino_reprocess_842.wav
|
| 384 |
+
./dataset/44k/chino_v7/chino_diff_aug_6_1.wav
|
| 385 |
+
./dataset/44k/chino_v7/chino_reprocess_734.wav
|
| 386 |
+
./dataset/44k/chino_v7/chino_diff_aug_273.wav
|
| 387 |
+
./dataset/44k/chino_v7/chino_diff_aug_338.wav
|
| 388 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_1.wav
|
| 389 |
+
./dataset/44k/chino_v7/chino_diff_aug_140_3.wav
|
| 390 |
+
./dataset/44k/chino_v7/chino_diff_aug_337.wav
|
| 391 |
+
./dataset/44k/chino_v7/chino_reprocess_837.wav
|
| 392 |
+
./dataset/44k/chino_v7/chino_diff_aug_229_0.wav
|
| 393 |
+
./dataset/44k/chino_v7/chino_diff_aug_139_2.wav
|
| 394 |
+
./dataset/44k/chino_v7/chino_reprocess_789.wav
|
| 395 |
+
./dataset/44k/chino_v7/chino_diff_aug_172.wav
|
| 396 |
+
./dataset/44k/chino_v7/chino_diff_aug_148_2.wav
|
| 397 |
+
./dataset/44k/chino_v7/chino_reprocess_818.wav
|
| 398 |
+
./dataset/44k/chino_v7/chino_diff_aug_116_5.wav
|
| 399 |
+
./dataset/44k/chino_v7/chino_diff_aug_69.wav
|
| 400 |
+
./dataset/44k/chino_v7/chino_reprocess_886.wav
|
| 401 |
+
./dataset/44k/chino_v7/chino_diff_aug_307.wav
|
| 402 |
+
./dataset/44k/chino_v7/chino_diff_aug_284.wav
|
| 403 |
+
./dataset/44k/chino_v7/chino_diff_aug_167_2.wav
|
| 404 |
+
./dataset/44k/chino_v7/chino_diff_aug_371.wav
|
| 405 |
+
./dataset/44k/chino_v7/chino_diff_aug_5.wav
|
| 406 |
+
./dataset/44k/chino_v7/chino_reprocess_813.wav
|
| 407 |
+
./dataset/44k/chino_v7/chino_diff_aug_131.wav
|
| 408 |
+
./dataset/44k/chino_v7/chino_diff_aug_166.wav
|
| 409 |
+
./dataset/44k/chino_v7/chino_diff_aug_154.wav
|
| 410 |
+
./dataset/44k/chino_v7/chino_diff_aug_279.wav
|
| 411 |
+
./dataset/44k/chino_v7/chino_diff_aug_380.wav
|
| 412 |
+
./dataset/44k/chino_v7/chino_diff_aug_204.wav
|
| 413 |
+
./dataset/44k/chino_v7/chino_diff_aug_174_1.wav
|
| 414 |
+
./dataset/44k/chino_v7/chino_diff_aug_360_1.wav
|
| 415 |
+
./dataset/44k/chino_v7/chino_reprocess_807.wav
|
| 416 |
+
./dataset/44k/chino_v7/chino_diff_aug_272.wav
|
| 417 |
+
./dataset/44k/chino_v7/chino_diff_aug_1_0.wav
|
| 418 |
+
./dataset/44k/chino_v7/chino_diff_aug_210.wav
|
| 419 |
+
./dataset/44k/chino_v7/chino_reprocess_749.wav
|
| 420 |
+
./dataset/44k/chino_v7/chino_diff_aug_347.wav
|
| 421 |
+
./dataset/44k/chino_v7/chino_diff_aug_349.wav
|
| 422 |
+
./dataset/44k/chino_v7/chino_diff_aug_344.wav
|
| 423 |
+
./dataset/44k/chino_v7/chino_reprocess_800.wav
|
| 424 |
+
./dataset/44k/chino_v7/chino_reprocess_769.wav
|
| 425 |
+
./dataset/44k/chino_v7/chino_reprocess_856.wav
|
| 426 |
+
./dataset/44k/chino_v7/chino_reprocess_808.wav
|
| 427 |
+
./dataset/44k/chino_v7/chino_diff_aug_197_2.wav
|
| 428 |
+
./dataset/44k/chino_v7/chino_diff_aug_295.wav
|
| 429 |
+
./dataset/44k/chino_v7/chino_diff_aug_169.wav
|
| 430 |
+
./dataset/44k/chino_v7/chino_diff_aug_174_2.wav
|
| 431 |
+
./dataset/44k/chino_v7/chino_reprocess_821.wav
|
| 432 |
+
./dataset/44k/chino_v7/chino_diff_aug_148_0.wav
|
| 433 |
+
./dataset/44k/chino_v7/chino_diff_aug_35.wav
|
| 434 |
+
./dataset/44k/chino_v7/chino_reprocess_820.wav
|
| 435 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_6.wav
|
| 436 |
+
./dataset/44k/chino_v7/chino_diff_aug_305.wav
|
| 437 |
+
./dataset/44k/chino_v7/chino_reprocess_811.wav
|
| 438 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_6.wav
|
| 439 |
+
./dataset/44k/chino_v7/chino_diff_aug_223.wav
|
| 440 |
+
./dataset/44k/chino_v7/chino_diff_aug_202_4.wav
|
| 441 |
+
./dataset/44k/chino_v7/chino_diff_aug_278.wav
|
| 442 |
+
./dataset/44k/chino_v7/chino_diff_aug_36_2.wav
|
| 443 |
+
./dataset/44k/chino_v7/chino_reprocess_834.wav
|
| 444 |
+
./dataset/44k/chino_v7/chino_diff_aug_196.wav
|
| 445 |
+
./dataset/44k/chino_v7/chino_reprocess_773.wav
|
| 446 |
+
./dataset/44k/chino_v7/chino_reprocess_719.wav
|
| 447 |
+
./dataset/44k/chino_v7/chino_diff_aug_157_0.wav
|
| 448 |
+
./dataset/44k/chino_v7/chino_diff_aug_1_1.wav
|
| 449 |
+
./dataset/44k/chino_v7/chino_reprocess_740.wav
|
| 450 |
+
./dataset/44k/chino_v7/chino_reprocess_815.wav
|
| 451 |
+
./dataset/44k/chino_v7/chino_diff_aug_180.wav
|
| 452 |
+
./dataset/44k/chino_v7/chino_reprocess_865.wav
|
| 453 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_0.wav
|
| 454 |
+
./dataset/44k/chino_v7/chino_diff_aug_179.wav
|
| 455 |
+
./dataset/44k/chino_v7/chino_diff_aug_140_2.wav
|
| 456 |
+
./dataset/44k/chino_v7/chino_reprocess_793.wav
|
| 457 |
+
./dataset/44k/chino_v7/chino_diff_aug_383.wav
|
| 458 |
+
./dataset/44k/chino_v7/chino_diff_aug_291.wav
|
| 459 |
+
./dataset/44k/chino_v7/chino_diff_aug_192.wav
|
| 460 |
+
./dataset/44k/chino_v7/chino_diff_aug_362.wav
|
| 461 |
+
./dataset/44k/chino_v7/chino_diff_aug_30.wav
|
| 462 |
+
./dataset/44k/chino_v7/chino_diff_aug_385.wav
|
| 463 |
+
./dataset/44k/chino_v7/chino_diff_aug_290.wav
|
| 464 |
+
./dataset/44k/chino_v7/chino_diff_aug_141.wav
|
| 465 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_5.wav
|
| 466 |
+
./dataset/44k/chino_v7/chino_diff_aug_287.wav
|
| 467 |
+
./dataset/44k/chino_v7/chino_diff_aug_33.wav
|
| 468 |
+
./dataset/44k/chino_v7/chino_diff_aug_186_2.wav
|
| 469 |
+
./dataset/44k/chino_v7/chino_diff_aug_366.wav
|
| 470 |
+
./dataset/44k/chino_v7/chino_diff_aug_213_0.wav
|
| 471 |
+
./dataset/44k/chino_v7/chino_diff_aug_259.wav
|
| 472 |
+
./dataset/44k/chino_v7/chino_diff_aug_170_1.wav
|
| 473 |
+
./dataset/44k/chino_v7/chino_diff_aug_47.wav
|
| 474 |
+
./dataset/44k/chino_v7/chino_reprocess_855.wav
|
| 475 |
+
./dataset/44k/chino_v7/chino_reprocess_863.wav
|
| 476 |
+
./dataset/44k/chino_v7/chino_diff_aug_391.wav
|
| 477 |
+
./dataset/44k/chino_v7/chino_diff_aug_110.wav
|
| 478 |
+
./dataset/44k/chino_v7/chino_reprocess_880.wav
|
| 479 |
+
./dataset/44k/chino_v7/chino_diff_aug_245.wav
|
| 480 |
+
./dataset/44k/chino_v7/chino_reprocess_754.wav
|
| 481 |
+
./dataset/44k/chino_v7/chino_diff_aug_209.wav
|
| 482 |
+
./dataset/44k/chino_v7/chino_diff_aug_227.wav
|
| 483 |
+
./dataset/44k/chino_v7/chino_diff_aug_401.wav
|
| 484 |
+
./dataset/44k/chino_v7/chino_diff_aug_353_2.wav
|
| 485 |
+
./dataset/44k/chino_v7/chino_diff_aug_167_3.wav
|
| 486 |
+
./dataset/44k/chino_v7/chino_reprocess_850.wav
|
| 487 |
+
./dataset/44k/chino_v7/chino_diff_aug_60.wav
|
| 488 |
+
./dataset/44k/chino_v7/chino_diff_aug_345.wav
|
| 489 |
+
./dataset/44k/chino_v7/chino_diff_aug_76.wav
|
| 490 |
+
./dataset/44k/chino_v7/chino_reprocess_832.wav
|
| 491 |
+
./dataset/44k/chino_v7/chino_diff_aug_155.wav
|
| 492 |
+
./dataset/44k/chino_v7/chino_reprocess_735.wav
|
| 493 |
+
./dataset/44k/chino_v7/chino_diff_aug_395.wav
|
| 494 |
+
./dataset/44k/chino_v7/chino_diff_aug_260.wav
|
| 495 |
+
./dataset/44k/chino_v7/chino_diff_aug_24.wav
|
| 496 |
+
./dataset/44k/chino_v7/chino_reprocess_763.wav
|
| 497 |
+
./dataset/44k/chino_v7/chino_diff_aug_16.wav
|
| 498 |
+
./dataset/44k/chino_v7/chino_diff_aug_107_3.wav
|
| 499 |
+
./dataset/44k/chino_v7/chino_diff_aug_232.wav
|
| 500 |
+
./dataset/44k/chino_v7/chino_diff_aug_405.wav
|
| 501 |
+
./dataset/44k/chino_v7/chino_reprocess_739.wav
|
| 502 |
+
./dataset/44k/chino_v7/chino_diff_aug_315_2.wav
|
| 503 |
+
./dataset/44k/chino_v7/chino_diff_aug_113_5.wav
|
| 504 |
+
./dataset/44k/chino_v7/chino_reprocess_876.wav
|
| 505 |
+
./dataset/44k/chino_v7/chino_diff_aug_343.wav
|
| 506 |
+
./dataset/44k/chino_v7/chino_reprocess_764.wav
|
| 507 |
+
./dataset/44k/chino_v7/chino_diff_aug_34.wav
|
| 508 |
+
./dataset/44k/chino_v7/chino_diff_aug_86.wav
|
| 509 |
+
./dataset/44k/chino_v7/chino_diff_aug_181.wav
|
| 510 |
+
./dataset/44k/chino_v7/chino_reprocess_760.wav
|
| 511 |
+
./dataset/44k/chino_v7/chino_diff_aug_197_0.wav
|
| 512 |
+
./dataset/44k/chino_v7/chino_diff_aug_163_2.wav
|
| 513 |
+
./dataset/44k/chino_v7/chino_diff_aug_390.wav
|
| 514 |
+
./dataset/44k/chino_v7/chino_reprocess_838.wav
|
| 515 |
+
./dataset/44k/chino_v7/chino_diff_aug_20.wav
|
| 516 |
+
./dataset/44k/chino_v7/chino_reprocess_748.wav
|
| 517 |
+
./dataset/44k/chino_v7/chino_reprocess_797.wav
|
| 518 |
+
./dataset/44k/chino_v7/chino_reprocess_768.wav
|
| 519 |
+
./dataset/44k/chino_v7/chino_diff_aug_216.wav
|
| 520 |
+
./dataset/44k/chino_v7/chino_diff_aug_331.wav
|
| 521 |
+
./dataset/44k/chino_v7/chino_reprocess_771.wav
|
| 522 |
+
./dataset/44k/chino_v7/chino_reprocess_745.wav
|
| 523 |
+
./dataset/44k/chino_v7/chino_diff_aug_83.wav
|
| 524 |
+
./dataset/44k/chino_v7/chino_diff_aug_178.wav
|
| 525 |
+
./dataset/44k/chino_v7/chino_reprocess_777.wav
|
| 526 |
+
./dataset/44k/chino_v7/chino_diff_aug_310_3.wav
|
| 527 |
+
./dataset/44k/chino_v7/chino_diff_aug_231_0.wav
|
| 528 |
+
./dataset/44k/chino_v7/chino_diff_aug_203.wav
|
| 529 |
+
./dataset/44k/chino_v7/chino_diff_aug_102.wav
|
| 530 |
+
./dataset/44k/chino_v7/chino_diff_aug_368.wav
|
| 531 |
+
./dataset/44k/chino_v7/chino_reprocess_783.wav
|
| 532 |
+
./dataset/44k/chino_v7/chino_diff_aug_346.wav
|
| 533 |
+
./dataset/44k/chino_v7/chino_diff_aug_70.wav
|
| 534 |
+
./dataset/44k/chino_v7/chino_diff_aug_25.wav
|
| 535 |
+
./dataset/44k/chino_v7/chino_diff_aug_389.wav
|
| 536 |
+
./dataset/44k/chino_v7/chino_diff_aug_190.wav
|
| 537 |
+
./dataset/44k/chino_v7/chino_reprocess_765.wav
|
| 538 |
+
./dataset/44k/chino_v7/chino_diff_aug_151_0.wav
|
| 539 |
+
./dataset/44k/chino_v7/chino_diff_aug_224.wav
|
| 540 |
+
./dataset/44k/chino_v7/chino_diff_aug_246.wav
|
| 541 |
+
./dataset/44k/chino_v7/chino_diff_aug_50.wav
|
| 542 |
+
./dataset/44k/chino_v7/chino_reprocess_753.wav
|
| 543 |
+
./dataset/44k/chino_v7/chino_diff_aug_56.wav
|
| 544 |
+
./dataset/44k/chino_v7/chino_reprocess_873.wav
|
| 545 |
+
./dataset/44k/chino_v7/chino_diff_aug_119.wav
|
| 546 |
+
./dataset/44k/chino_v7/chino_reprocess_884.wav
|
| 547 |
+
./dataset/44k/chino_v7/chino_diff_aug_21.wav
|
| 548 |
+
./dataset/44k/chino_v7/chino_diff_aug_280.wav
|
| 549 |
+
./dataset/44k/chino_v7/chino_diff_aug_218.wav
|
| 550 |
+
./dataset/44k/chino_v7/chino_reprocess_883.wav
|
| 551 |
+
./dataset/44k/chino_v7/chino_reprocess_782.wav
|
| 552 |
+
./dataset/44k/chino_v7/chino_diff_aug_350.wav
|
| 553 |
+
./dataset/44k/chino_v7/chino_reprocess_847.wav
|
| 554 |
+
./dataset/44k/chino_v7/chino_diff_aug_193_4.wav
|
| 555 |
+
./dataset/44k/chino_v7/chino_diff_aug_173.wav
|
| 556 |
+
./dataset/44k/chino_v7/chino_diff_aug_44.wav
|
| 557 |
+
./dataset/44k/chino_v7/chino_diff_aug_255.wav
|
| 558 |
+
./dataset/44k/chino_v7/chino_diff_aug_376.wav
|
| 559 |
+
./dataset/44k/chino_v7/chino_diff_aug_199.wav
|
| 560 |
+
./dataset/44k/chino_v7/chino_diff_aug_105.wav
|
| 561 |
+
./dataset/44k/chino_v7/chino_diff_aug_149_0.wav
|
| 562 |
+
./dataset/44k/chino_v7/chino_diff_aug_136.wav
|
| 563 |
+
./dataset/44k/chino_v7/chino_diff_aug_335.wav
|
| 564 |
+
./dataset/44k/chino_v7/chino_reprocess_824.wav
|
| 565 |
+
./dataset/44k/chino_v7/chino_diff_aug_1_2.wav
|
| 566 |
+
./dataset/44k/chino_v7/chino_diff_aug_8_0.wav
|
filelists/val.txt
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
./dataset/44k/chino_v7/chino_diff_aug_197_1.wav
|
| 2 |
+
./dataset/44k/chino_v7/chino_diff_aug_237_3.wav
|
inference/__init__.py
ADDED
|
File without changes
|
inference/__pycache__/__init__.cpython-38.pyc
ADDED
|
Binary file (159 Bytes). View file
|
|
|
inference/__pycache__/infer_tool.cpython-38.pyc
ADDED
|
Binary file (14.8 kB). View file
|
|
|
inference/__pycache__/infer_tool_webui.cpython-38.pyc
ADDED
|
Binary file (15.3 kB). View file
|
|
|
inference/__pycache__/slicer.cpython-38.pyc
ADDED
|
Binary file (3.86 kB). View file
|
|
|
inference/chunks_temp.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"info": "temp_dict"}
|
inference/infer_tool.py
ADDED
|
@@ -0,0 +1,546 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gc
|
| 2 |
+
import hashlib
|
| 3 |
+
import io
|
| 4 |
+
import json
|
| 5 |
+
import logging
|
| 6 |
+
import os
|
| 7 |
+
import pickle
|
| 8 |
+
import time
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
|
| 11 |
+
import librosa
|
| 12 |
+
import numpy as np
|
| 13 |
+
|
| 14 |
+
# import onnxruntime
|
| 15 |
+
import soundfile
|
| 16 |
+
import torch
|
| 17 |
+
import torchaudio
|
| 18 |
+
|
| 19 |
+
import cluster
|
| 20 |
+
import utils
|
| 21 |
+
from diffusion.unit2mel import load_model_vocoder
|
| 22 |
+
from inference import slicer
|
| 23 |
+
from models import SynthesizerTrn
|
| 24 |
+
|
| 25 |
+
logging.getLogger('matplotlib').setLevel(logging.WARNING)
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def read_temp(file_name):
|
| 29 |
+
if not os.path.exists(file_name):
|
| 30 |
+
with open(file_name, "w") as f:
|
| 31 |
+
f.write(json.dumps({"info": "temp_dict"}))
|
| 32 |
+
return {}
|
| 33 |
+
else:
|
| 34 |
+
try:
|
| 35 |
+
with open(file_name, "r") as f:
|
| 36 |
+
data = f.read()
|
| 37 |
+
data_dict = json.loads(data)
|
| 38 |
+
if os.path.getsize(file_name) > 50 * 1024 * 1024:
|
| 39 |
+
f_name = file_name.replace("\\", "/").split("/")[-1]
|
| 40 |
+
print(f"clean {f_name}")
|
| 41 |
+
for wav_hash in list(data_dict.keys()):
|
| 42 |
+
if int(time.time()) - int(data_dict[wav_hash]["time"]) > 14 * 24 * 3600:
|
| 43 |
+
del data_dict[wav_hash]
|
| 44 |
+
except Exception as e:
|
| 45 |
+
print(e)
|
| 46 |
+
print(f"{file_name} error,auto rebuild file")
|
| 47 |
+
data_dict = {"info": "temp_dict"}
|
| 48 |
+
return data_dict
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def write_temp(file_name, data):
|
| 52 |
+
with open(file_name, "w") as f:
|
| 53 |
+
f.write(json.dumps(data))
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def timeit(func):
|
| 57 |
+
def run(*args, **kwargs):
|
| 58 |
+
t = time.time()
|
| 59 |
+
res = func(*args, **kwargs)
|
| 60 |
+
print('executing \'%s\' costed %.3fs' % (func.__name__, time.time() - t))
|
| 61 |
+
return res
|
| 62 |
+
|
| 63 |
+
return run
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
def format_wav(audio_path):
|
| 67 |
+
if Path(audio_path).suffix == '.wav':
|
| 68 |
+
return
|
| 69 |
+
raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True, sr=None)
|
| 70 |
+
soundfile.write(Path(audio_path).with_suffix(".wav"), raw_audio, raw_sample_rate)
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
def get_end_file(dir_path, end):
|
| 74 |
+
file_lists = []
|
| 75 |
+
for root, dirs, files in os.walk(dir_path):
|
| 76 |
+
files = [f for f in files if f[0] != '.']
|
| 77 |
+
dirs[:] = [d for d in dirs if d[0] != '.']
|
| 78 |
+
for f_file in files:
|
| 79 |
+
if f_file.endswith(end):
|
| 80 |
+
file_lists.append(os.path.join(root, f_file).replace("\\", "/"))
|
| 81 |
+
return file_lists
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def get_md5(content):
|
| 85 |
+
return hashlib.new("md5", content).hexdigest()
|
| 86 |
+
|
| 87 |
+
def fill_a_to_b(a, b):
|
| 88 |
+
if len(a) < len(b):
|
| 89 |
+
for _ in range(0, len(b) - len(a)):
|
| 90 |
+
a.append(a[0])
|
| 91 |
+
|
| 92 |
+
def mkdir(paths: list):
|
| 93 |
+
for path in paths:
|
| 94 |
+
if not os.path.exists(path):
|
| 95 |
+
os.mkdir(path)
|
| 96 |
+
|
| 97 |
+
def pad_array(arr, target_length):
|
| 98 |
+
current_length = arr.shape[0]
|
| 99 |
+
if current_length >= target_length:
|
| 100 |
+
return arr
|
| 101 |
+
else:
|
| 102 |
+
pad_width = target_length - current_length
|
| 103 |
+
pad_left = pad_width // 2
|
| 104 |
+
pad_right = pad_width - pad_left
|
| 105 |
+
padded_arr = np.pad(arr, (pad_left, pad_right), 'constant', constant_values=(0, 0))
|
| 106 |
+
return padded_arr
|
| 107 |
+
|
| 108 |
+
def split_list_by_n(list_collection, n, pre=0):
|
| 109 |
+
for i in range(0, len(list_collection), n):
|
| 110 |
+
yield list_collection[i-pre if i-pre>=0 else i: i + n]
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
class F0FilterException(Exception):
|
| 114 |
+
pass
|
| 115 |
+
|
| 116 |
+
class Svc(object):
|
| 117 |
+
def __init__(self, net_g_path, config_path,
|
| 118 |
+
device=None,
|
| 119 |
+
cluster_model_path="logs/44k/kmeans_10000.pt",
|
| 120 |
+
nsf_hifigan_enhance = False,
|
| 121 |
+
diffusion_model_path="logs/44k/diffusion/model_0.pt",
|
| 122 |
+
diffusion_config_path="configs/diffusion.yaml",
|
| 123 |
+
shallow_diffusion = False,
|
| 124 |
+
only_diffusion = False,
|
| 125 |
+
spk_mix_enable = False,
|
| 126 |
+
feature_retrieval = False
|
| 127 |
+
):
|
| 128 |
+
self.net_g_path = net_g_path
|
| 129 |
+
self.only_diffusion = only_diffusion
|
| 130 |
+
self.shallow_diffusion = shallow_diffusion
|
| 131 |
+
self.feature_retrieval = feature_retrieval
|
| 132 |
+
if device is None:
|
| 133 |
+
self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 134 |
+
else:
|
| 135 |
+
self.dev = torch.device(device)
|
| 136 |
+
self.net_g_ms = None
|
| 137 |
+
if not self.only_diffusion:
|
| 138 |
+
self.hps_ms = utils.get_hparams_from_file(config_path,True)
|
| 139 |
+
self.target_sample = self.hps_ms.data.sampling_rate
|
| 140 |
+
self.hop_size = self.hps_ms.data.hop_length
|
| 141 |
+
self.spk2id = self.hps_ms.spk
|
| 142 |
+
self.unit_interpolate_mode = self.hps_ms.data.unit_interpolate_mode if self.hps_ms.data.unit_interpolate_mode is not None else 'left'
|
| 143 |
+
self.vol_embedding = self.hps_ms.model.vol_embedding if self.hps_ms.model.vol_embedding is not None else False
|
| 144 |
+
self.speech_encoder = self.hps_ms.model.speech_encoder if self.hps_ms.model.speech_encoder is not None else 'vec768l12'
|
| 145 |
+
|
| 146 |
+
self.nsf_hifigan_enhance = nsf_hifigan_enhance
|
| 147 |
+
if self.shallow_diffusion or self.only_diffusion:
|
| 148 |
+
if os.path.exists(diffusion_model_path) and os.path.exists(diffusion_model_path):
|
| 149 |
+
self.diffusion_model,self.vocoder,self.diffusion_args = load_model_vocoder(diffusion_model_path,self.dev,config_path=diffusion_config_path)
|
| 150 |
+
if self.only_diffusion:
|
| 151 |
+
self.target_sample = self.diffusion_args.data.sampling_rate
|
| 152 |
+
self.hop_size = self.diffusion_args.data.block_size
|
| 153 |
+
self.spk2id = self.diffusion_args.spk
|
| 154 |
+
self.dtype = torch.float32
|
| 155 |
+
self.speech_encoder = self.diffusion_args.data.encoder
|
| 156 |
+
self.unit_interpolate_mode = self.diffusion_args.data.unit_interpolate_mode if self.diffusion_args.data.unit_interpolate_mode is not None else 'left'
|
| 157 |
+
if spk_mix_enable:
|
| 158 |
+
self.diffusion_model.init_spkmix(len(self.spk2id))
|
| 159 |
+
else:
|
| 160 |
+
print("No diffusion model or config found. Shallow diffusion mode will False")
|
| 161 |
+
self.shallow_diffusion = self.only_diffusion = False
|
| 162 |
+
|
| 163 |
+
# load hubert and model
|
| 164 |
+
if not self.only_diffusion:
|
| 165 |
+
self.load_model(spk_mix_enable)
|
| 166 |
+
self.hubert_model = utils.get_speech_encoder(self.speech_encoder,device=self.dev)
|
| 167 |
+
self.volume_extractor = utils.Volume_Extractor(self.hop_size)
|
| 168 |
+
else:
|
| 169 |
+
self.hubert_model = utils.get_speech_encoder(self.diffusion_args.data.encoder,device=self.dev)
|
| 170 |
+
self.volume_extractor = utils.Volume_Extractor(self.diffusion_args.data.block_size)
|
| 171 |
+
|
| 172 |
+
if os.path.exists(cluster_model_path):
|
| 173 |
+
if self.feature_retrieval:
|
| 174 |
+
with open(cluster_model_path,"rb") as f:
|
| 175 |
+
self.cluster_model = pickle.load(f)
|
| 176 |
+
self.big_npy = None
|
| 177 |
+
self.now_spk_id = -1
|
| 178 |
+
else:
|
| 179 |
+
self.cluster_model = cluster.get_cluster_model(cluster_model_path)
|
| 180 |
+
else:
|
| 181 |
+
self.feature_retrieval=False
|
| 182 |
+
|
| 183 |
+
if self.shallow_diffusion :
|
| 184 |
+
self.nsf_hifigan_enhance = False
|
| 185 |
+
if self.nsf_hifigan_enhance:
|
| 186 |
+
from modules.enhancer import Enhancer
|
| 187 |
+
self.enhancer = Enhancer('nsf-hifigan', 'pretrain/nsf_hifigan/model',device=self.dev)
|
| 188 |
+
|
| 189 |
+
def load_model(self, spk_mix_enable=False):
|
| 190 |
+
# get model configuration
|
| 191 |
+
self.net_g_ms = SynthesizerTrn(
|
| 192 |
+
self.hps_ms.data.filter_length // 2 + 1,
|
| 193 |
+
self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
|
| 194 |
+
**self.hps_ms.model)
|
| 195 |
+
_ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
|
| 196 |
+
self.dtype = list(self.net_g_ms.parameters())[0].dtype
|
| 197 |
+
if "half" in self.net_g_path and torch.cuda.is_available():
|
| 198 |
+
_ = self.net_g_ms.half().eval().to(self.dev)
|
| 199 |
+
else:
|
| 200 |
+
_ = self.net_g_ms.eval().to(self.dev)
|
| 201 |
+
if spk_mix_enable:
|
| 202 |
+
self.net_g_ms.EnableCharacterMix(len(self.spk2id), self.dev)
|
| 203 |
+
|
| 204 |
+
def get_unit_f0(self, wav, tran, cluster_infer_ratio, speaker, f0_filter ,f0_predictor,cr_threshold=0.05):
|
| 205 |
+
|
| 206 |
+
if not hasattr(self,"f0_predictor_object") or self.f0_predictor_object is None or f0_predictor != self.f0_predictor_object.name:
|
| 207 |
+
self.f0_predictor_object = utils.get_f0_predictor(f0_predictor,hop_length=self.hop_size,sampling_rate=self.target_sample,device=self.dev,threshold=cr_threshold)
|
| 208 |
+
f0, uv = self.f0_predictor_object.compute_f0_uv(wav)
|
| 209 |
+
|
| 210 |
+
if f0_filter and sum(f0) == 0:
|
| 211 |
+
raise F0FilterException("No voice detected")
|
| 212 |
+
f0 = torch.FloatTensor(f0).to(self.dev)
|
| 213 |
+
uv = torch.FloatTensor(uv).to(self.dev)
|
| 214 |
+
|
| 215 |
+
f0 = f0 * 2 ** (tran / 12)
|
| 216 |
+
f0 = f0.unsqueeze(0)
|
| 217 |
+
uv = uv.unsqueeze(0)
|
| 218 |
+
|
| 219 |
+
wav = torch.from_numpy(wav).to(self.dev)
|
| 220 |
+
if not hasattr(self,"audio16k_resample_transform"):
|
| 221 |
+
self.audio16k_resample_transform = torchaudio.transforms.Resample(self.target_sample, 16000).to(self.dev)
|
| 222 |
+
wav16k = self.audio16k_resample_transform(wav[None,:])[0]
|
| 223 |
+
|
| 224 |
+
c = self.hubert_model.encoder(wav16k)
|
| 225 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1],self.unit_interpolate_mode)
|
| 226 |
+
|
| 227 |
+
if cluster_infer_ratio !=0:
|
| 228 |
+
if self.feature_retrieval:
|
| 229 |
+
speaker_id = self.spk2id.get(speaker)
|
| 230 |
+
if not speaker_id and type(speaker) is int:
|
| 231 |
+
if len(self.spk2id.__dict__) >= speaker:
|
| 232 |
+
speaker_id = speaker
|
| 233 |
+
if speaker_id is None:
|
| 234 |
+
raise RuntimeError("The name you entered is not in the speaker list!")
|
| 235 |
+
feature_index = self.cluster_model[speaker_id]
|
| 236 |
+
feat_np = np.ascontiguousarray(c.transpose(0,1).cpu().numpy())
|
| 237 |
+
if self.big_npy is None or self.now_spk_id != speaker_id:
|
| 238 |
+
self.big_npy = feature_index.reconstruct_n(0, feature_index.ntotal)
|
| 239 |
+
self.now_spk_id = speaker_id
|
| 240 |
+
print("starting feature retrieval...")
|
| 241 |
+
score, ix = feature_index.search(feat_np, k=8)
|
| 242 |
+
weight = np.square(1 / score)
|
| 243 |
+
weight /= weight.sum(axis=1, keepdims=True)
|
| 244 |
+
npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
|
| 245 |
+
c = cluster_infer_ratio * npy + (1 - cluster_infer_ratio) * feat_np
|
| 246 |
+
c = torch.FloatTensor(c).to(self.dev).transpose(0,1)
|
| 247 |
+
print("end feature retrieval...")
|
| 248 |
+
else:
|
| 249 |
+
cluster_c = cluster.get_cluster_center_result(self.cluster_model, c.cpu().numpy().T, speaker).T
|
| 250 |
+
cluster_c = torch.FloatTensor(cluster_c).to(self.dev)
|
| 251 |
+
c = cluster_infer_ratio * cluster_c + (1 - cluster_infer_ratio) * c
|
| 252 |
+
|
| 253 |
+
c = c.unsqueeze(0)
|
| 254 |
+
return c, f0, uv
|
| 255 |
+
|
| 256 |
+
def infer(self, speaker, tran, raw_path,
|
| 257 |
+
cluster_infer_ratio=0,
|
| 258 |
+
auto_predict_f0=False,
|
| 259 |
+
noice_scale=0.4,
|
| 260 |
+
f0_filter=False,
|
| 261 |
+
f0_predictor='pm',
|
| 262 |
+
enhancer_adaptive_key = 0,
|
| 263 |
+
cr_threshold = 0.05,
|
| 264 |
+
k_step = 100,
|
| 265 |
+
frame = 0,
|
| 266 |
+
spk_mix = False,
|
| 267 |
+
second_encoding = False,
|
| 268 |
+
loudness_envelope_adjustment = 1
|
| 269 |
+
):
|
| 270 |
+
torchaudio.set_audio_backend("soundfile")
|
| 271 |
+
wav, sr = torchaudio.load(raw_path)
|
| 272 |
+
if not hasattr(self,"audio_resample_transform") or self.audio16k_resample_transform.orig_freq != sr:
|
| 273 |
+
self.audio_resample_transform = torchaudio.transforms.Resample(sr,self.target_sample)
|
| 274 |
+
wav = self.audio_resample_transform(wav).numpy()[0]
|
| 275 |
+
if spk_mix:
|
| 276 |
+
c, f0, uv = self.get_unit_f0(wav, tran, 0, None, f0_filter,f0_predictor,cr_threshold=cr_threshold)
|
| 277 |
+
n_frames = f0.size(1)
|
| 278 |
+
sid = speaker[:, frame:frame+n_frames].transpose(0,1)
|
| 279 |
+
else:
|
| 280 |
+
speaker_id = self.spk2id.get(speaker)
|
| 281 |
+
if not speaker_id and type(speaker) is int:
|
| 282 |
+
if len(self.spk2id.__dict__) >= speaker:
|
| 283 |
+
speaker_id = speaker
|
| 284 |
+
if speaker_id is None:
|
| 285 |
+
raise RuntimeError("The name you entered is not in the speaker list!")
|
| 286 |
+
sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
|
| 287 |
+
c, f0, uv = self.get_unit_f0(wav, tran, cluster_infer_ratio, speaker, f0_filter,f0_predictor,cr_threshold=cr_threshold)
|
| 288 |
+
n_frames = f0.size(1)
|
| 289 |
+
c = c.to(self.dtype)
|
| 290 |
+
f0 = f0.to(self.dtype)
|
| 291 |
+
uv = uv.to(self.dtype)
|
| 292 |
+
with torch.no_grad():
|
| 293 |
+
start = time.time()
|
| 294 |
+
vol = None
|
| 295 |
+
if not self.only_diffusion:
|
| 296 |
+
vol = self.volume_extractor.extract(torch.FloatTensor(wav).to(self.dev)[None,:])[None,:].to(self.dev) if self.vol_embedding else None
|
| 297 |
+
audio,f0 = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale,vol=vol)
|
| 298 |
+
audio = audio[0,0].data.float()
|
| 299 |
+
audio_mel = self.vocoder.extract(audio[None,:],self.target_sample) if self.shallow_diffusion else None
|
| 300 |
+
else:
|
| 301 |
+
audio = torch.FloatTensor(wav).to(self.dev)
|
| 302 |
+
audio_mel = None
|
| 303 |
+
if self.dtype != torch.float32:
|
| 304 |
+
c = c.to(torch.float32)
|
| 305 |
+
f0 = f0.to(torch.float32)
|
| 306 |
+
uv = uv.to(torch.float32)
|
| 307 |
+
if self.only_diffusion or self.shallow_diffusion:
|
| 308 |
+
vol = self.volume_extractor.extract(audio[None,:])[None,:,None].to(self.dev) if vol is None else vol[:,:,None]
|
| 309 |
+
if self.shallow_diffusion and second_encoding:
|
| 310 |
+
if not hasattr(self,"audio16k_resample_transform"):
|
| 311 |
+
self.audio16k_resample_transform = torchaudio.transforms.Resample(self.target_sample, 16000).to(self.dev)
|
| 312 |
+
audio16k = self.audio16k_resample_transform(audio[None,:])[0]
|
| 313 |
+
c = self.hubert_model.encoder(audio16k)
|
| 314 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1],self.unit_interpolate_mode)
|
| 315 |
+
f0 = f0[:,:,None]
|
| 316 |
+
c = c.transpose(-1,-2)
|
| 317 |
+
audio_mel = self.diffusion_model(
|
| 318 |
+
c,
|
| 319 |
+
f0,
|
| 320 |
+
vol,
|
| 321 |
+
spk_id = sid,
|
| 322 |
+
spk_mix_dict = None,
|
| 323 |
+
gt_spec=audio_mel,
|
| 324 |
+
infer=True,
|
| 325 |
+
infer_speedup=self.diffusion_args.infer.speedup,
|
| 326 |
+
method=self.diffusion_args.infer.method,
|
| 327 |
+
k_step=k_step)
|
| 328 |
+
audio = self.vocoder.infer(audio_mel, f0).squeeze()
|
| 329 |
+
if self.nsf_hifigan_enhance:
|
| 330 |
+
audio, _ = self.enhancer.enhance(
|
| 331 |
+
audio[None,:],
|
| 332 |
+
self.target_sample,
|
| 333 |
+
f0[:,:,None],
|
| 334 |
+
self.hps_ms.data.hop_length,
|
| 335 |
+
adaptive_key = enhancer_adaptive_key)
|
| 336 |
+
if loudness_envelope_adjustment != 1:
|
| 337 |
+
audio = utils.change_rms(wav,self.target_sample,audio,self.target_sample,loudness_envelope_adjustment)
|
| 338 |
+
use_time = time.time() - start
|
| 339 |
+
print("vits use time:{}".format(use_time))
|
| 340 |
+
return audio, audio.shape[-1], n_frames
|
| 341 |
+
|
| 342 |
+
def clear_empty(self):
|
| 343 |
+
# clean up vram
|
| 344 |
+
torch.cuda.empty_cache()
|
| 345 |
+
|
| 346 |
+
def unload_model(self):
|
| 347 |
+
# unload model
|
| 348 |
+
self.net_g_ms = self.net_g_ms.to("cpu")
|
| 349 |
+
del self.net_g_ms
|
| 350 |
+
if hasattr(self,"enhancer"):
|
| 351 |
+
self.enhancer.enhancer = self.enhancer.enhancer.to("cpu")
|
| 352 |
+
del self.enhancer.enhancer
|
| 353 |
+
del self.enhancer
|
| 354 |
+
gc.collect()
|
| 355 |
+
|
| 356 |
+
def slice_inference(self,
|
| 357 |
+
raw_audio_path,
|
| 358 |
+
spk,
|
| 359 |
+
tran,
|
| 360 |
+
slice_db,
|
| 361 |
+
cluster_infer_ratio,
|
| 362 |
+
auto_predict_f0,
|
| 363 |
+
noice_scale,
|
| 364 |
+
pad_seconds=0.5,
|
| 365 |
+
clip_seconds=0,
|
| 366 |
+
lg_num=0,
|
| 367 |
+
lgr_num =0.75,
|
| 368 |
+
f0_predictor='pm',
|
| 369 |
+
enhancer_adaptive_key = 0,
|
| 370 |
+
cr_threshold = 0.05,
|
| 371 |
+
k_step = 100,
|
| 372 |
+
use_spk_mix = False,
|
| 373 |
+
second_encoding = False,
|
| 374 |
+
loudness_envelope_adjustment = 1
|
| 375 |
+
):
|
| 376 |
+
if use_spk_mix:
|
| 377 |
+
if len(self.spk2id) == 1:
|
| 378 |
+
spk = self.spk2id.keys()[0]
|
| 379 |
+
use_spk_mix = False
|
| 380 |
+
wav_path = Path(raw_audio_path).with_suffix('.wav')
|
| 381 |
+
chunks = slicer.cut(wav_path, db_thresh=slice_db)
|
| 382 |
+
audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
|
| 383 |
+
per_size = int(clip_seconds*audio_sr)
|
| 384 |
+
lg_size = int(lg_num*audio_sr)
|
| 385 |
+
lg_size_r = int(lg_size*lgr_num)
|
| 386 |
+
lg_size_c_l = (lg_size-lg_size_r)//2
|
| 387 |
+
lg_size_c_r = lg_size-lg_size_r-lg_size_c_l
|
| 388 |
+
lg = np.linspace(0,1,lg_size_r) if lg_size!=0 else 0
|
| 389 |
+
|
| 390 |
+
if use_spk_mix:
|
| 391 |
+
assert len(self.spk2id) == len(spk)
|
| 392 |
+
audio_length = 0
|
| 393 |
+
for (slice_tag, data) in audio_data:
|
| 394 |
+
aud_length = int(np.ceil(len(data) / audio_sr * self.target_sample))
|
| 395 |
+
if slice_tag:
|
| 396 |
+
audio_length += aud_length // self.hop_size
|
| 397 |
+
continue
|
| 398 |
+
if per_size != 0:
|
| 399 |
+
datas = split_list_by_n(data, per_size,lg_size)
|
| 400 |
+
else:
|
| 401 |
+
datas = [data]
|
| 402 |
+
for k,dat in enumerate(datas):
|
| 403 |
+
pad_len = int(audio_sr * pad_seconds)
|
| 404 |
+
per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample))
|
| 405 |
+
a_length = per_length + 2 * pad_len
|
| 406 |
+
audio_length += a_length // self.hop_size
|
| 407 |
+
audio_length += len(audio_data)
|
| 408 |
+
spk_mix_tensor = torch.zeros(size=(len(spk), audio_length)).to(self.dev)
|
| 409 |
+
for i in range(len(spk)):
|
| 410 |
+
last_end = None
|
| 411 |
+
for mix in spk[i]:
|
| 412 |
+
if mix[3]<0. or mix[2]<0.:
|
| 413 |
+
raise RuntimeError("mix value must higer Than zero!")
|
| 414 |
+
begin = int(audio_length * mix[0])
|
| 415 |
+
end = int(audio_length * mix[1])
|
| 416 |
+
length = end - begin
|
| 417 |
+
if length<=0:
|
| 418 |
+
raise RuntimeError("begin Must lower Than end!")
|
| 419 |
+
step = (mix[3] - mix[2])/length
|
| 420 |
+
if last_end is not None:
|
| 421 |
+
if last_end != begin:
|
| 422 |
+
raise RuntimeError("[i]EndTime Must Equal [i+1]BeginTime!")
|
| 423 |
+
last_end = end
|
| 424 |
+
if step == 0.:
|
| 425 |
+
spk_mix_data = torch.zeros(length).to(self.dev) + mix[2]
|
| 426 |
+
else:
|
| 427 |
+
spk_mix_data = torch.arange(mix[2],mix[3],step).to(self.dev)
|
| 428 |
+
if(len(spk_mix_data)<length):
|
| 429 |
+
num_pad = length - len(spk_mix_data)
|
| 430 |
+
spk_mix_data = torch.nn.functional.pad(spk_mix_data, [0, num_pad], mode="reflect").to(self.dev)
|
| 431 |
+
spk_mix_tensor[i][begin:end] = spk_mix_data[:length]
|
| 432 |
+
|
| 433 |
+
spk_mix_ten = torch.sum(spk_mix_tensor,dim=0).unsqueeze(0).to(self.dev)
|
| 434 |
+
# spk_mix_tensor[0][spk_mix_ten<0.001] = 1.0
|
| 435 |
+
for i, x in enumerate(spk_mix_ten[0]):
|
| 436 |
+
if x == 0.0:
|
| 437 |
+
spk_mix_ten[0][i] = 1.0
|
| 438 |
+
spk_mix_tensor[:,i] = 1.0 / len(spk)
|
| 439 |
+
spk_mix_tensor = spk_mix_tensor / spk_mix_ten
|
| 440 |
+
if not ((torch.sum(spk_mix_tensor,dim=0) - 1.)<0.0001).all():
|
| 441 |
+
raise RuntimeError("sum(spk_mix_tensor) not equal 1")
|
| 442 |
+
spk = spk_mix_tensor
|
| 443 |
+
|
| 444 |
+
global_frame = 0
|
| 445 |
+
audio = []
|
| 446 |
+
for (slice_tag, data) in audio_data:
|
| 447 |
+
print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
|
| 448 |
+
# padd
|
| 449 |
+
length = int(np.ceil(len(data) / audio_sr * self.target_sample))
|
| 450 |
+
if slice_tag:
|
| 451 |
+
print('jump empty segment')
|
| 452 |
+
_audio = np.zeros(length)
|
| 453 |
+
audio.extend(list(pad_array(_audio, length)))
|
| 454 |
+
global_frame += length // self.hop_size
|
| 455 |
+
continue
|
| 456 |
+
if per_size != 0:
|
| 457 |
+
datas = split_list_by_n(data, per_size,lg_size)
|
| 458 |
+
else:
|
| 459 |
+
datas = [data]
|
| 460 |
+
for k,dat in enumerate(datas):
|
| 461 |
+
per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample)) if clip_seconds!=0 else length
|
| 462 |
+
if clip_seconds!=0:
|
| 463 |
+
print(f'###=====segment clip start, {round(len(dat) / audio_sr, 3)}s======')
|
| 464 |
+
# padd
|
| 465 |
+
pad_len = int(audio_sr * pad_seconds)
|
| 466 |
+
dat = np.concatenate([np.zeros([pad_len]), dat, np.zeros([pad_len])])
|
| 467 |
+
raw_path = io.BytesIO()
|
| 468 |
+
soundfile.write(raw_path, dat, audio_sr, format="wav")
|
| 469 |
+
raw_path.seek(0)
|
| 470 |
+
out_audio, out_sr, out_frame = self.infer(spk, tran, raw_path,
|
| 471 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
| 472 |
+
auto_predict_f0=auto_predict_f0,
|
| 473 |
+
noice_scale=noice_scale,
|
| 474 |
+
f0_predictor = f0_predictor,
|
| 475 |
+
enhancer_adaptive_key = enhancer_adaptive_key,
|
| 476 |
+
cr_threshold = cr_threshold,
|
| 477 |
+
k_step = k_step,
|
| 478 |
+
frame = global_frame,
|
| 479 |
+
spk_mix = use_spk_mix,
|
| 480 |
+
second_encoding = second_encoding,
|
| 481 |
+
loudness_envelope_adjustment = loudness_envelope_adjustment
|
| 482 |
+
)
|
| 483 |
+
global_frame += out_frame
|
| 484 |
+
_audio = out_audio.cpu().numpy()
|
| 485 |
+
pad_len = int(self.target_sample * pad_seconds)
|
| 486 |
+
_audio = _audio[pad_len:-pad_len]
|
| 487 |
+
_audio = pad_array(_audio, per_length)
|
| 488 |
+
if lg_size!=0 and k!=0:
|
| 489 |
+
lg1 = audio[-(lg_size_r+lg_size_c_r):-lg_size_c_r] if lgr_num != 1 else audio[-lg_size:]
|
| 490 |
+
lg2 = _audio[lg_size_c_l:lg_size_c_l+lg_size_r] if lgr_num != 1 else _audio[0:lg_size]
|
| 491 |
+
lg_pre = lg1*(1-lg)+lg2*lg
|
| 492 |
+
audio = audio[0:-(lg_size_r+lg_size_c_r)] if lgr_num != 1 else audio[0:-lg_size]
|
| 493 |
+
audio.extend(lg_pre)
|
| 494 |
+
_audio = _audio[lg_size_c_l+lg_size_r:] if lgr_num != 1 else _audio[lg_size:]
|
| 495 |
+
audio.extend(list(_audio))
|
| 496 |
+
return np.array(audio)
|
| 497 |
+
|
| 498 |
+
class RealTimeVC:
|
| 499 |
+
def __init__(self):
|
| 500 |
+
self.last_chunk = None
|
| 501 |
+
self.last_o = None
|
| 502 |
+
self.chunk_len = 16000 # chunk length
|
| 503 |
+
self.pre_len = 3840 # cross fade length, multiples of 640
|
| 504 |
+
|
| 505 |
+
# Input and output are 1-dimensional numpy waveform arrays
|
| 506 |
+
|
| 507 |
+
def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path,
|
| 508 |
+
cluster_infer_ratio=0,
|
| 509 |
+
auto_predict_f0=False,
|
| 510 |
+
noice_scale=0.4,
|
| 511 |
+
f0_filter=False):
|
| 512 |
+
|
| 513 |
+
import maad
|
| 514 |
+
audio, sr = torchaudio.load(input_wav_path)
|
| 515 |
+
audio = audio.cpu().numpy()[0]
|
| 516 |
+
temp_wav = io.BytesIO()
|
| 517 |
+
if self.last_chunk is None:
|
| 518 |
+
input_wav_path.seek(0)
|
| 519 |
+
|
| 520 |
+
audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path,
|
| 521 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
| 522 |
+
auto_predict_f0=auto_predict_f0,
|
| 523 |
+
noice_scale=noice_scale,
|
| 524 |
+
f0_filter=f0_filter)
|
| 525 |
+
|
| 526 |
+
audio = audio.cpu().numpy()
|
| 527 |
+
self.last_chunk = audio[-self.pre_len:]
|
| 528 |
+
self.last_o = audio
|
| 529 |
+
return audio[-self.chunk_len:]
|
| 530 |
+
else:
|
| 531 |
+
audio = np.concatenate([self.last_chunk, audio])
|
| 532 |
+
soundfile.write(temp_wav, audio, sr, format="wav")
|
| 533 |
+
temp_wav.seek(0)
|
| 534 |
+
|
| 535 |
+
audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav,
|
| 536 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
| 537 |
+
auto_predict_f0=auto_predict_f0,
|
| 538 |
+
noice_scale=noice_scale,
|
| 539 |
+
f0_filter=f0_filter)
|
| 540 |
+
|
| 541 |
+
audio = audio.cpu().numpy()
|
| 542 |
+
ret = maad.util.crossfade(self.last_o, audio, self.pre_len)
|
| 543 |
+
self.last_chunk = audio[-self.pre_len:]
|
| 544 |
+
self.last_o = audio
|
| 545 |
+
return ret[self.chunk_len:2 * self.chunk_len]
|
| 546 |
+
|
inference/infer_tool_grad.py
ADDED
|
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import io
|
| 2 |
+
import logging
|
| 3 |
+
import os
|
| 4 |
+
|
| 5 |
+
import librosa
|
| 6 |
+
import numpy as np
|
| 7 |
+
import parselmouth
|
| 8 |
+
import soundfile
|
| 9 |
+
import torch
|
| 10 |
+
import torchaudio
|
| 11 |
+
|
| 12 |
+
import utils
|
| 13 |
+
from inference import slicer
|
| 14 |
+
from models import SynthesizerTrn
|
| 15 |
+
|
| 16 |
+
logging.getLogger('numba').setLevel(logging.WARNING)
|
| 17 |
+
logging.getLogger('matplotlib').setLevel(logging.WARNING)
|
| 18 |
+
|
| 19 |
+
def resize2d_f0(x, target_len):
|
| 20 |
+
source = np.array(x)
|
| 21 |
+
source[source < 0.001] = np.nan
|
| 22 |
+
target = np.interp(np.arange(0, len(source) * target_len, len(source)) / target_len, np.arange(0, len(source)),
|
| 23 |
+
source)
|
| 24 |
+
res = np.nan_to_num(target)
|
| 25 |
+
return res
|
| 26 |
+
|
| 27 |
+
def get_f0(x, p_len,f0_up_key=0):
|
| 28 |
+
|
| 29 |
+
time_step = 160 / 16000 * 1000
|
| 30 |
+
f0_min = 50
|
| 31 |
+
f0_max = 1100
|
| 32 |
+
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
|
| 33 |
+
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
|
| 34 |
+
|
| 35 |
+
f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
|
| 36 |
+
time_step=time_step / 1000, voicing_threshold=0.6,
|
| 37 |
+
pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
|
| 38 |
+
|
| 39 |
+
pad_size=(p_len - len(f0) + 1) // 2
|
| 40 |
+
if(pad_size>0 or p_len - len(f0) - pad_size>0):
|
| 41 |
+
f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
|
| 42 |
+
|
| 43 |
+
f0 *= pow(2, f0_up_key / 12)
|
| 44 |
+
f0_mel = 1127 * np.log(1 + f0 / 700)
|
| 45 |
+
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
|
| 46 |
+
f0_mel[f0_mel <= 1] = 1
|
| 47 |
+
f0_mel[f0_mel > 255] = 255
|
| 48 |
+
f0_coarse = np.rint(f0_mel).astype(np.int)
|
| 49 |
+
return f0_coarse, f0
|
| 50 |
+
|
| 51 |
+
def clean_pitch(input_pitch):
|
| 52 |
+
num_nan = np.sum(input_pitch == 1)
|
| 53 |
+
if num_nan / len(input_pitch) > 0.9:
|
| 54 |
+
input_pitch[input_pitch != 1] = 1
|
| 55 |
+
return input_pitch
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def plt_pitch(input_pitch):
|
| 59 |
+
input_pitch = input_pitch.astype(float)
|
| 60 |
+
input_pitch[input_pitch == 1] = np.nan
|
| 61 |
+
return input_pitch
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
def f0_to_pitch(ff):
|
| 65 |
+
f0_pitch = 69 + 12 * np.log2(ff / 440)
|
| 66 |
+
return f0_pitch
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def fill_a_to_b(a, b):
|
| 70 |
+
if len(a) < len(b):
|
| 71 |
+
for _ in range(0, len(b) - len(a)):
|
| 72 |
+
a.append(a[0])
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def mkdir(paths: list):
|
| 76 |
+
for path in paths:
|
| 77 |
+
if not os.path.exists(path):
|
| 78 |
+
os.mkdir(path)
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
class VitsSvc(object):
|
| 82 |
+
def __init__(self):
|
| 83 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 84 |
+
self.SVCVITS = None
|
| 85 |
+
self.hps = None
|
| 86 |
+
self.speakers = None
|
| 87 |
+
self.hubert_soft = utils.get_hubert_model()
|
| 88 |
+
|
| 89 |
+
def set_device(self, device):
|
| 90 |
+
self.device = torch.device(device)
|
| 91 |
+
self.hubert_soft.to(self.device)
|
| 92 |
+
if self.SVCVITS is not None:
|
| 93 |
+
self.SVCVITS.to(self.device)
|
| 94 |
+
|
| 95 |
+
def loadCheckpoint(self, path):
|
| 96 |
+
self.hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
|
| 97 |
+
self.SVCVITS = SynthesizerTrn(
|
| 98 |
+
self.hps.data.filter_length // 2 + 1,
|
| 99 |
+
self.hps.train.segment_size // self.hps.data.hop_length,
|
| 100 |
+
**self.hps.model)
|
| 101 |
+
_ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", self.SVCVITS, None)
|
| 102 |
+
_ = self.SVCVITS.eval().to(self.device)
|
| 103 |
+
self.speakers = self.hps.spk
|
| 104 |
+
|
| 105 |
+
def get_units(self, source, sr):
|
| 106 |
+
source = source.unsqueeze(0).to(self.device)
|
| 107 |
+
with torch.inference_mode():
|
| 108 |
+
units = self.hubert_soft.units(source)
|
| 109 |
+
return units
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def get_unit_pitch(self, in_path, tran):
|
| 113 |
+
source, sr = torchaudio.load(in_path)
|
| 114 |
+
source = torchaudio.functional.resample(source, sr, 16000)
|
| 115 |
+
if len(source.shape) == 2 and source.shape[1] >= 2:
|
| 116 |
+
source = torch.mean(source, dim=0).unsqueeze(0)
|
| 117 |
+
soft = self.get_units(source, sr).squeeze(0).cpu().numpy()
|
| 118 |
+
f0_coarse, f0 = get_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran)
|
| 119 |
+
return soft, f0
|
| 120 |
+
|
| 121 |
+
def infer(self, speaker_id, tran, raw_path):
|
| 122 |
+
speaker_id = self.speakers[speaker_id]
|
| 123 |
+
sid = torch.LongTensor([int(speaker_id)]).to(self.device).unsqueeze(0)
|
| 124 |
+
soft, pitch = self.get_unit_pitch(raw_path, tran)
|
| 125 |
+
f0 = torch.FloatTensor(clean_pitch(pitch)).unsqueeze(0).to(self.device)
|
| 126 |
+
stn_tst = torch.FloatTensor(soft)
|
| 127 |
+
with torch.no_grad():
|
| 128 |
+
x_tst = stn_tst.unsqueeze(0).to(self.device)
|
| 129 |
+
x_tst = torch.repeat_interleave(x_tst, repeats=2, dim=1).transpose(1, 2)
|
| 130 |
+
audio,_ = self.SVCVITS.infer(x_tst, f0=f0, g=sid)[0,0].data.float()
|
| 131 |
+
return audio, audio.shape[-1]
|
| 132 |
+
|
| 133 |
+
def inference(self,srcaudio,chara,tran,slice_db):
|
| 134 |
+
sampling_rate, audio = srcaudio
|
| 135 |
+
audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
|
| 136 |
+
if len(audio.shape) > 1:
|
| 137 |
+
audio = librosa.to_mono(audio.transpose(1, 0))
|
| 138 |
+
if sampling_rate != 16000:
|
| 139 |
+
audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
|
| 140 |
+
soundfile.write("tmpwav.wav", audio, 16000, format="wav")
|
| 141 |
+
chunks = slicer.cut("tmpwav.wav", db_thresh=slice_db)
|
| 142 |
+
audio_data, audio_sr = slicer.chunks2audio("tmpwav.wav", chunks)
|
| 143 |
+
audio = []
|
| 144 |
+
for (slice_tag, data) in audio_data:
|
| 145 |
+
length = int(np.ceil(len(data) / audio_sr * self.hps.data.sampling_rate))
|
| 146 |
+
raw_path = io.BytesIO()
|
| 147 |
+
soundfile.write(raw_path, data, audio_sr, format="wav")
|
| 148 |
+
raw_path.seek(0)
|
| 149 |
+
if slice_tag:
|
| 150 |
+
_audio = np.zeros(length)
|
| 151 |
+
else:
|
| 152 |
+
out_audio, out_sr = self.infer(chara, tran, raw_path)
|
| 153 |
+
_audio = out_audio.cpu().numpy()
|
| 154 |
+
audio.extend(list(_audio))
|
| 155 |
+
audio = (np.array(audio) * 32768.0).astype('int16')
|
| 156 |
+
return (self.hps.data.sampling_rate,audio)
|
inference/infer_tool_webui.py
ADDED
|
@@ -0,0 +1,547 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gc
|
| 2 |
+
import hashlib
|
| 3 |
+
import io
|
| 4 |
+
import json
|
| 5 |
+
import logging
|
| 6 |
+
import os
|
| 7 |
+
import pickle
|
| 8 |
+
import time
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
|
| 11 |
+
import librosa
|
| 12 |
+
import numpy as np
|
| 13 |
+
|
| 14 |
+
# import onnxruntime
|
| 15 |
+
import soundfile
|
| 16 |
+
import torch
|
| 17 |
+
import torchaudio
|
| 18 |
+
from tqdm import tqdm
|
| 19 |
+
|
| 20 |
+
import cluster
|
| 21 |
+
import utils
|
| 22 |
+
from diffusion.unit2mel import load_model_vocoder
|
| 23 |
+
from inference import slicer
|
| 24 |
+
from models import SynthesizerTrn
|
| 25 |
+
|
| 26 |
+
logging.getLogger('matplotlib').setLevel(logging.WARNING)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def read_temp(file_name):
|
| 30 |
+
if not os.path.exists(file_name):
|
| 31 |
+
with open(file_name, "w") as f:
|
| 32 |
+
f.write(json.dumps({"info": "temp_dict"}))
|
| 33 |
+
return {}
|
| 34 |
+
else:
|
| 35 |
+
try:
|
| 36 |
+
with open(file_name, "r") as f:
|
| 37 |
+
data = f.read()
|
| 38 |
+
data_dict = json.loads(data)
|
| 39 |
+
if os.path.getsize(file_name) > 50 * 1024 * 1024:
|
| 40 |
+
f_name = file_name.replace("\\", "/").split("/")[-1]
|
| 41 |
+
print(f"clean {f_name}")
|
| 42 |
+
for wav_hash in list(data_dict.keys()):
|
| 43 |
+
if int(time.time()) - int(data_dict[wav_hash]["time"]) > 14 * 24 * 3600:
|
| 44 |
+
del data_dict[wav_hash]
|
| 45 |
+
except Exception as e:
|
| 46 |
+
print(e)
|
| 47 |
+
print(f"{file_name} error,auto rebuild file")
|
| 48 |
+
data_dict = {"info": "temp_dict"}
|
| 49 |
+
return data_dict
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
def write_temp(file_name, data):
|
| 53 |
+
with open(file_name, "w") as f:
|
| 54 |
+
f.write(json.dumps(data))
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def timeit(func):
|
| 58 |
+
def run(*args, **kwargs):
|
| 59 |
+
t = time.time()
|
| 60 |
+
res = func(*args, **kwargs)
|
| 61 |
+
print('executing \'%s\' costed %.3fs' % (func.__name__, time.time() - t))
|
| 62 |
+
return res
|
| 63 |
+
|
| 64 |
+
return run
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def format_wav(audio_path):
|
| 68 |
+
if Path(audio_path).suffix == '.wav':
|
| 69 |
+
return
|
| 70 |
+
raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True, sr=None)
|
| 71 |
+
soundfile.write(Path(audio_path).with_suffix(".wav"), raw_audio, raw_sample_rate)
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def get_end_file(dir_path, end):
|
| 75 |
+
file_lists = []
|
| 76 |
+
for root, dirs, files in os.walk(dir_path):
|
| 77 |
+
files = [f for f in files if f[0] != '.']
|
| 78 |
+
dirs[:] = [d for d in dirs if d[0] != '.']
|
| 79 |
+
for f_file in files:
|
| 80 |
+
if f_file.endswith(end):
|
| 81 |
+
file_lists.append(os.path.join(root, f_file).replace("\\", "/"))
|
| 82 |
+
return file_lists
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
def get_md5(content):
|
| 86 |
+
return hashlib.new("md5", content).hexdigest()
|
| 87 |
+
|
| 88 |
+
def fill_a_to_b(a, b):
|
| 89 |
+
if len(a) < len(b):
|
| 90 |
+
for _ in range(0, len(b) - len(a)):
|
| 91 |
+
a.append(a[0])
|
| 92 |
+
|
| 93 |
+
def mkdir(paths: list):
|
| 94 |
+
for path in paths:
|
| 95 |
+
if not os.path.exists(path):
|
| 96 |
+
os.mkdir(path)
|
| 97 |
+
|
| 98 |
+
def pad_array(arr, target_length):
|
| 99 |
+
current_length = arr.shape[0]
|
| 100 |
+
if current_length >= target_length:
|
| 101 |
+
return arr
|
| 102 |
+
else:
|
| 103 |
+
pad_width = target_length - current_length
|
| 104 |
+
pad_left = pad_width // 2
|
| 105 |
+
pad_right = pad_width - pad_left
|
| 106 |
+
padded_arr = np.pad(arr, (pad_left, pad_right), 'constant', constant_values=(0, 0))
|
| 107 |
+
return padded_arr
|
| 108 |
+
|
| 109 |
+
def split_list_by_n(list_collection, n, pre=0):
|
| 110 |
+
for i in range(0, len(list_collection), n):
|
| 111 |
+
yield list_collection[i-pre if i-pre>=0 else i: i + n]
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
class F0FilterException(Exception):
|
| 115 |
+
pass
|
| 116 |
+
|
| 117 |
+
class Svc(object):
|
| 118 |
+
def __init__(self, net_g_path, config_path,
|
| 119 |
+
device=None,
|
| 120 |
+
cluster_model_path="logs/44k/kmeans_10000.pt",
|
| 121 |
+
nsf_hifigan_enhance = False,
|
| 122 |
+
diffusion_model_path="logs/44k/diffusion/model_0.pt",
|
| 123 |
+
diffusion_config_path="configs/diffusion.yaml",
|
| 124 |
+
shallow_diffusion = False,
|
| 125 |
+
only_diffusion = False,
|
| 126 |
+
spk_mix_enable = False,
|
| 127 |
+
feature_retrieval = False
|
| 128 |
+
):
|
| 129 |
+
self.net_g_path = net_g_path
|
| 130 |
+
self.only_diffusion = only_diffusion
|
| 131 |
+
self.shallow_diffusion = shallow_diffusion
|
| 132 |
+
self.feature_retrieval = feature_retrieval
|
| 133 |
+
if device is None:
|
| 134 |
+
self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 135 |
+
else:
|
| 136 |
+
self.dev = torch.device(device)
|
| 137 |
+
self.net_g_ms = None
|
| 138 |
+
if not self.only_diffusion:
|
| 139 |
+
self.hps_ms = utils.get_hparams_from_file(config_path,True)
|
| 140 |
+
self.target_sample = self.hps_ms.data.sampling_rate
|
| 141 |
+
self.hop_size = self.hps_ms.data.hop_length
|
| 142 |
+
self.spk2id = self.hps_ms.spk
|
| 143 |
+
self.unit_interpolate_mode = self.hps_ms.data.unit_interpolate_mode if self.hps_ms.data.unit_interpolate_mode is not None else 'left'
|
| 144 |
+
self.vol_embedding = self.hps_ms.model.vol_embedding if self.hps_ms.model.vol_embedding is not None else False
|
| 145 |
+
self.speech_encoder = self.hps_ms.model.speech_encoder if self.hps_ms.model.speech_encoder is not None else 'vec768l12'
|
| 146 |
+
|
| 147 |
+
self.nsf_hifigan_enhance = nsf_hifigan_enhance
|
| 148 |
+
if self.shallow_diffusion or self.only_diffusion:
|
| 149 |
+
if os.path.exists(diffusion_model_path) and os.path.exists(diffusion_model_path):
|
| 150 |
+
self.diffusion_model,self.vocoder,self.diffusion_args = load_model_vocoder(diffusion_model_path,self.dev,config_path=diffusion_config_path)
|
| 151 |
+
if self.only_diffusion:
|
| 152 |
+
self.target_sample = self.diffusion_args.data.sampling_rate
|
| 153 |
+
self.hop_size = self.diffusion_args.data.block_size
|
| 154 |
+
self.spk2id = self.diffusion_args.spk
|
| 155 |
+
self.dtype = torch.float32
|
| 156 |
+
self.speech_encoder = self.diffusion_args.data.encoder
|
| 157 |
+
self.unit_interpolate_mode = self.diffusion_args.data.unit_interpolate_mode if self.diffusion_args.data.unit_interpolate_mode is not None else 'left'
|
| 158 |
+
if spk_mix_enable:
|
| 159 |
+
self.diffusion_model.init_spkmix(len(self.spk2id))
|
| 160 |
+
else:
|
| 161 |
+
print("No diffusion model or config found. Shallow diffusion mode will False")
|
| 162 |
+
self.shallow_diffusion = self.only_diffusion = False
|
| 163 |
+
|
| 164 |
+
# load hubert and model
|
| 165 |
+
if not self.only_diffusion:
|
| 166 |
+
self.load_model(spk_mix_enable)
|
| 167 |
+
self.hubert_model = utils.get_speech_encoder(self.speech_encoder,device=self.dev)
|
| 168 |
+
self.volume_extractor = utils.Volume_Extractor(self.hop_size)
|
| 169 |
+
else:
|
| 170 |
+
self.hubert_model = utils.get_speech_encoder(self.diffusion_args.data.encoder,device=self.dev)
|
| 171 |
+
self.volume_extractor = utils.Volume_Extractor(self.diffusion_args.data.block_size)
|
| 172 |
+
|
| 173 |
+
if os.path.exists(cluster_model_path):
|
| 174 |
+
if self.feature_retrieval:
|
| 175 |
+
with open(cluster_model_path,"rb") as f:
|
| 176 |
+
self.cluster_model = pickle.load(f)
|
| 177 |
+
self.big_npy = None
|
| 178 |
+
self.now_spk_id = -1
|
| 179 |
+
else:
|
| 180 |
+
self.cluster_model = cluster.get_cluster_model(cluster_model_path)
|
| 181 |
+
else:
|
| 182 |
+
self.feature_retrieval=False
|
| 183 |
+
|
| 184 |
+
if self.shallow_diffusion :
|
| 185 |
+
self.nsf_hifigan_enhance = False
|
| 186 |
+
if self.nsf_hifigan_enhance:
|
| 187 |
+
from modules.enhancer import Enhancer
|
| 188 |
+
self.enhancer = Enhancer('nsf-hifigan', 'pretrain/nsf_hifigan/model',device=self.dev)
|
| 189 |
+
|
| 190 |
+
def load_model(self, spk_mix_enable=False):
|
| 191 |
+
# get model configuration
|
| 192 |
+
self.net_g_ms = SynthesizerTrn(
|
| 193 |
+
self.hps_ms.data.filter_length // 2 + 1,
|
| 194 |
+
self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
|
| 195 |
+
**self.hps_ms.model)
|
| 196 |
+
_ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
|
| 197 |
+
self.dtype = list(self.net_g_ms.parameters())[0].dtype
|
| 198 |
+
if "half" in self.net_g_path and torch.cuda.is_available():
|
| 199 |
+
_ = self.net_g_ms.half().eval().to(self.dev)
|
| 200 |
+
else:
|
| 201 |
+
_ = self.net_g_ms.eval().to(self.dev)
|
| 202 |
+
if spk_mix_enable:
|
| 203 |
+
self.net_g_ms.EnableCharacterMix(len(self.spk2id), self.dev)
|
| 204 |
+
|
| 205 |
+
def get_unit_f0(self, wav, tran, cluster_infer_ratio, speaker, f0_filter ,f0_predictor,cr_threshold=0.05):
|
| 206 |
+
|
| 207 |
+
if not hasattr(self,"f0_predictor_object") or self.f0_predictor_object is None or f0_predictor != self.f0_predictor_object.name:
|
| 208 |
+
self.f0_predictor_object = utils.get_f0_predictor(f0_predictor,hop_length=self.hop_size,sampling_rate=self.target_sample,device=self.dev,threshold=cr_threshold)
|
| 209 |
+
f0, uv = self.f0_predictor_object.compute_f0_uv(wav)
|
| 210 |
+
|
| 211 |
+
if f0_filter and sum(f0) == 0:
|
| 212 |
+
raise F0FilterException("No voice detected")
|
| 213 |
+
f0 = torch.FloatTensor(f0).to(self.dev)
|
| 214 |
+
uv = torch.FloatTensor(uv).to(self.dev)
|
| 215 |
+
|
| 216 |
+
f0 = f0 * 2 ** (tran / 12)
|
| 217 |
+
f0 = f0.unsqueeze(0)
|
| 218 |
+
uv = uv.unsqueeze(0)
|
| 219 |
+
|
| 220 |
+
wav = torch.from_numpy(wav).to(self.dev)
|
| 221 |
+
if not hasattr(self,"audio16k_resample_transform"):
|
| 222 |
+
self.audio16k_resample_transform = torchaudio.transforms.Resample(self.target_sample, 16000).to(self.dev)
|
| 223 |
+
wav16k = self.audio16k_resample_transform(wav[None,:])[0]
|
| 224 |
+
|
| 225 |
+
c = self.hubert_model.encoder(wav16k)
|
| 226 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1],self.unit_interpolate_mode)
|
| 227 |
+
|
| 228 |
+
if cluster_infer_ratio !=0:
|
| 229 |
+
if self.feature_retrieval:
|
| 230 |
+
speaker_id = self.spk2id.get(speaker)
|
| 231 |
+
if not speaker_id and type(speaker) is int:
|
| 232 |
+
if len(self.spk2id.__dict__) >= speaker:
|
| 233 |
+
speaker_id = speaker
|
| 234 |
+
if speaker_id is None:
|
| 235 |
+
raise RuntimeError("The name you entered is not in the speaker list!")
|
| 236 |
+
feature_index = self.cluster_model[speaker_id]
|
| 237 |
+
feat_np = np.ascontiguousarray(c.transpose(0,1).cpu().numpy())
|
| 238 |
+
if self.big_npy is None or self.now_spk_id != speaker_id:
|
| 239 |
+
self.big_npy = feature_index.reconstruct_n(0, feature_index.ntotal)
|
| 240 |
+
self.now_spk_id = speaker_id
|
| 241 |
+
print("starting feature retrieval...")
|
| 242 |
+
score, ix = feature_index.search(feat_np, k=8)
|
| 243 |
+
weight = np.square(1 / score)
|
| 244 |
+
weight /= weight.sum(axis=1, keepdims=True)
|
| 245 |
+
npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
|
| 246 |
+
c = cluster_infer_ratio * npy + (1 - cluster_infer_ratio) * feat_np
|
| 247 |
+
c = torch.FloatTensor(c).to(self.dev).transpose(0,1)
|
| 248 |
+
print("end feature retrieval...")
|
| 249 |
+
else:
|
| 250 |
+
cluster_c = cluster.get_cluster_center_result(self.cluster_model, c.cpu().numpy().T, speaker).T
|
| 251 |
+
cluster_c = torch.FloatTensor(cluster_c).to(self.dev)
|
| 252 |
+
c = cluster_infer_ratio * cluster_c + (1 - cluster_infer_ratio) * c
|
| 253 |
+
|
| 254 |
+
c = c.unsqueeze(0)
|
| 255 |
+
return c, f0, uv
|
| 256 |
+
|
| 257 |
+
def infer(self, speaker, tran, raw_path,
|
| 258 |
+
cluster_infer_ratio=0,
|
| 259 |
+
auto_predict_f0=False,
|
| 260 |
+
noice_scale=0.4,
|
| 261 |
+
f0_filter=False,
|
| 262 |
+
f0_predictor='pm',
|
| 263 |
+
enhancer_adaptive_key = 0,
|
| 264 |
+
cr_threshold = 0.05,
|
| 265 |
+
k_step = 100,
|
| 266 |
+
frame = 0,
|
| 267 |
+
spk_mix = False,
|
| 268 |
+
second_encoding = False,
|
| 269 |
+
loudness_envelope_adjustment = 1
|
| 270 |
+
):
|
| 271 |
+
torchaudio.set_audio_backend("soundfile")
|
| 272 |
+
wav, sr = torchaudio.load(raw_path)
|
| 273 |
+
if not hasattr(self,"audio_resample_transform") or self.audio16k_resample_transform.orig_freq != sr:
|
| 274 |
+
self.audio_resample_transform = torchaudio.transforms.Resample(sr,self.target_sample)
|
| 275 |
+
wav = self.audio_resample_transform(wav).numpy()[0]
|
| 276 |
+
if spk_mix:
|
| 277 |
+
c, f0, uv = self.get_unit_f0(wav, tran, 0, None, f0_filter,f0_predictor,cr_threshold=cr_threshold)
|
| 278 |
+
n_frames = f0.size(1)
|
| 279 |
+
sid = speaker[:, frame:frame+n_frames].transpose(0,1)
|
| 280 |
+
else:
|
| 281 |
+
speaker_id = self.spk2id.get(speaker)
|
| 282 |
+
if not speaker_id and type(speaker) is int:
|
| 283 |
+
if len(self.spk2id.__dict__) >= speaker:
|
| 284 |
+
speaker_id = speaker
|
| 285 |
+
if speaker_id is None:
|
| 286 |
+
raise RuntimeError("The name you entered is not in the speaker list!")
|
| 287 |
+
sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
|
| 288 |
+
c, f0, uv = self.get_unit_f0(wav, tran, cluster_infer_ratio, speaker, f0_filter,f0_predictor,cr_threshold=cr_threshold)
|
| 289 |
+
n_frames = f0.size(1)
|
| 290 |
+
c = c.to(self.dtype)
|
| 291 |
+
f0 = f0.to(self.dtype)
|
| 292 |
+
uv = uv.to(self.dtype)
|
| 293 |
+
with torch.no_grad():
|
| 294 |
+
start = time.time()
|
| 295 |
+
vol = None
|
| 296 |
+
if not self.only_diffusion:
|
| 297 |
+
vol = self.volume_extractor.extract(torch.FloatTensor(wav).to(self.dev)[None,:])[None,:].to(self.dev) if self.vol_embedding else None
|
| 298 |
+
audio,f0 = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale,vol=vol)
|
| 299 |
+
audio = audio[0,0].data.float()
|
| 300 |
+
audio_mel = self.vocoder.extract(audio[None,:],self.target_sample) if self.shallow_diffusion else None
|
| 301 |
+
else:
|
| 302 |
+
audio = torch.FloatTensor(wav).to(self.dev)
|
| 303 |
+
audio_mel = None
|
| 304 |
+
if self.dtype != torch.float32:
|
| 305 |
+
c = c.to(torch.float32)
|
| 306 |
+
f0 = f0.to(torch.float32)
|
| 307 |
+
uv = uv.to(torch.float32)
|
| 308 |
+
if self.only_diffusion or self.shallow_diffusion:
|
| 309 |
+
vol = self.volume_extractor.extract(audio[None,:])[None,:,None].to(self.dev) if vol is None else vol[:,:,None]
|
| 310 |
+
if self.shallow_diffusion and second_encoding:
|
| 311 |
+
if not hasattr(self,"audio16k_resample_transform"):
|
| 312 |
+
self.audio16k_resample_transform = torchaudio.transforms.Resample(self.target_sample, 16000).to(self.dev)
|
| 313 |
+
audio16k = self.audio16k_resample_transform(audio[None,:])[0]
|
| 314 |
+
c = self.hubert_model.encoder(audio16k)
|
| 315 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1],self.unit_interpolate_mode)
|
| 316 |
+
f0 = f0[:,:,None]
|
| 317 |
+
c = c.transpose(-1,-2)
|
| 318 |
+
audio_mel = self.diffusion_model(
|
| 319 |
+
c,
|
| 320 |
+
f0,
|
| 321 |
+
vol,
|
| 322 |
+
spk_id = sid,
|
| 323 |
+
spk_mix_dict = None,
|
| 324 |
+
gt_spec=audio_mel,
|
| 325 |
+
infer=True,
|
| 326 |
+
infer_speedup=self.diffusion_args.infer.speedup,
|
| 327 |
+
method=self.diffusion_args.infer.method,
|
| 328 |
+
k_step=k_step)
|
| 329 |
+
audio = self.vocoder.infer(audio_mel, f0).squeeze()
|
| 330 |
+
if self.nsf_hifigan_enhance:
|
| 331 |
+
audio, _ = self.enhancer.enhance(
|
| 332 |
+
audio[None,:],
|
| 333 |
+
self.target_sample,
|
| 334 |
+
f0[:,:,None],
|
| 335 |
+
self.hps_ms.data.hop_length,
|
| 336 |
+
adaptive_key = enhancer_adaptive_key)
|
| 337 |
+
if loudness_envelope_adjustment != 1:
|
| 338 |
+
audio = utils.change_rms(wav,self.target_sample,audio,self.target_sample,loudness_envelope_adjustment)
|
| 339 |
+
use_time = time.time() - start
|
| 340 |
+
print("vits use time:{}".format(use_time))
|
| 341 |
+
return audio, audio.shape[-1], n_frames
|
| 342 |
+
|
| 343 |
+
def clear_empty(self):
|
| 344 |
+
# clean up vram
|
| 345 |
+
torch.cuda.empty_cache()
|
| 346 |
+
|
| 347 |
+
def unload_model(self):
|
| 348 |
+
# unload model
|
| 349 |
+
self.net_g_ms = self.net_g_ms.to("cpu")
|
| 350 |
+
del self.net_g_ms
|
| 351 |
+
if hasattr(self,"enhancer"):
|
| 352 |
+
self.enhancer.enhancer = self.enhancer.enhancer.to("cpu")
|
| 353 |
+
del self.enhancer.enhancer
|
| 354 |
+
del self.enhancer
|
| 355 |
+
gc.collect()
|
| 356 |
+
|
| 357 |
+
def slice_inference(self,
|
| 358 |
+
raw_audio_path,
|
| 359 |
+
spk,
|
| 360 |
+
tran,
|
| 361 |
+
slice_db,
|
| 362 |
+
cluster_infer_ratio,
|
| 363 |
+
auto_predict_f0,
|
| 364 |
+
noice_scale,
|
| 365 |
+
pad_seconds=0.5,
|
| 366 |
+
clip_seconds=0,
|
| 367 |
+
lg_num=0,
|
| 368 |
+
lgr_num =0.75,
|
| 369 |
+
f0_predictor='pm',
|
| 370 |
+
enhancer_adaptive_key = 0,
|
| 371 |
+
cr_threshold = 0.05,
|
| 372 |
+
k_step = 100,
|
| 373 |
+
use_spk_mix = False,
|
| 374 |
+
second_encoding = False,
|
| 375 |
+
loudness_envelope_adjustment = 1
|
| 376 |
+
):
|
| 377 |
+
if use_spk_mix:
|
| 378 |
+
if len(self.spk2id) == 1:
|
| 379 |
+
spk = self.spk2id.keys()[0]
|
| 380 |
+
use_spk_mix = False
|
| 381 |
+
wav_path = Path(raw_audio_path).with_suffix('.wav')
|
| 382 |
+
chunks = slicer.cut(wav_path, db_thresh=slice_db)
|
| 383 |
+
audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
|
| 384 |
+
per_size = int(clip_seconds*audio_sr)
|
| 385 |
+
lg_size = int(lg_num*audio_sr)
|
| 386 |
+
lg_size_r = int(lg_size*lgr_num)
|
| 387 |
+
lg_size_c_l = (lg_size-lg_size_r)//2
|
| 388 |
+
lg_size_c_r = lg_size-lg_size_r-lg_size_c_l
|
| 389 |
+
lg = np.linspace(0,1,lg_size_r) if lg_size!=0 else 0
|
| 390 |
+
|
| 391 |
+
if use_spk_mix:
|
| 392 |
+
assert len(self.spk2id) == len(spk)
|
| 393 |
+
audio_length = 0
|
| 394 |
+
for (slice_tag, data) in audio_data:
|
| 395 |
+
aud_length = int(np.ceil(len(data) / audio_sr * self.target_sample))
|
| 396 |
+
if slice_tag:
|
| 397 |
+
audio_length += aud_length // self.hop_size
|
| 398 |
+
continue
|
| 399 |
+
if per_size != 0:
|
| 400 |
+
datas = split_list_by_n(data, per_size,lg_size)
|
| 401 |
+
else:
|
| 402 |
+
datas = [data]
|
| 403 |
+
for k,dat in enumerate(datas):
|
| 404 |
+
pad_len = int(audio_sr * pad_seconds)
|
| 405 |
+
per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample))
|
| 406 |
+
a_length = per_length + 2 * pad_len
|
| 407 |
+
audio_length += a_length // self.hop_size
|
| 408 |
+
audio_length += len(audio_data)
|
| 409 |
+
spk_mix_tensor = torch.zeros(size=(len(spk), audio_length)).to(self.dev)
|
| 410 |
+
for i in range(len(spk)):
|
| 411 |
+
last_end = None
|
| 412 |
+
for mix in spk[i]:
|
| 413 |
+
if mix[3]<0. or mix[2]<0.:
|
| 414 |
+
raise RuntimeError("mix value must higer Than zero!")
|
| 415 |
+
begin = int(audio_length * mix[0])
|
| 416 |
+
end = int(audio_length * mix[1])
|
| 417 |
+
length = end - begin
|
| 418 |
+
if length<=0:
|
| 419 |
+
raise RuntimeError("begin Must lower Than end!")
|
| 420 |
+
step = (mix[3] - mix[2])/length
|
| 421 |
+
if last_end is not None:
|
| 422 |
+
if last_end != begin:
|
| 423 |
+
raise RuntimeError("[i]EndTime Must Equal [i+1]BeginTime!")
|
| 424 |
+
last_end = end
|
| 425 |
+
if step == 0.:
|
| 426 |
+
spk_mix_data = torch.zeros(length).to(self.dev) + mix[2]
|
| 427 |
+
else:
|
| 428 |
+
spk_mix_data = torch.arange(mix[2],mix[3],step).to(self.dev)
|
| 429 |
+
if(len(spk_mix_data)<length):
|
| 430 |
+
num_pad = length - len(spk_mix_data)
|
| 431 |
+
spk_mix_data = torch.nn.functional.pad(spk_mix_data, [0, num_pad], mode="reflect").to(self.dev)
|
| 432 |
+
spk_mix_tensor[i][begin:end] = spk_mix_data[:length]
|
| 433 |
+
|
| 434 |
+
spk_mix_ten = torch.sum(spk_mix_tensor,dim=0).unsqueeze(0).to(self.dev)
|
| 435 |
+
# spk_mix_tensor[0][spk_mix_ten<0.001] = 1.0
|
| 436 |
+
for i, x in enumerate(spk_mix_ten[0]):
|
| 437 |
+
if x == 0.0:
|
| 438 |
+
spk_mix_ten[0][i] = 1.0
|
| 439 |
+
spk_mix_tensor[:,i] = 1.0 / len(spk)
|
| 440 |
+
spk_mix_tensor = spk_mix_tensor / spk_mix_ten
|
| 441 |
+
if not ((torch.sum(spk_mix_tensor,dim=0) - 1.)<0.0001).all():
|
| 442 |
+
raise RuntimeError("sum(spk_mix_tensor) not equal 1")
|
| 443 |
+
spk = spk_mix_tensor
|
| 444 |
+
|
| 445 |
+
global_frame = 0
|
| 446 |
+
audio = []
|
| 447 |
+
for (slice_tag, data) in tqdm(audio_data):
|
| 448 |
+
print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
|
| 449 |
+
# padd
|
| 450 |
+
length = int(np.ceil(len(data) / audio_sr * self.target_sample))
|
| 451 |
+
if slice_tag:
|
| 452 |
+
print('jump empty segment')
|
| 453 |
+
_audio = np.zeros(length)
|
| 454 |
+
audio.extend(list(pad_array(_audio, length)))
|
| 455 |
+
global_frame += length // self.hop_size
|
| 456 |
+
continue
|
| 457 |
+
if per_size != 0:
|
| 458 |
+
datas = split_list_by_n(data, per_size,lg_size)
|
| 459 |
+
else:
|
| 460 |
+
datas = [data]
|
| 461 |
+
for k,dat in enumerate(datas):
|
| 462 |
+
per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample)) if clip_seconds!=0 else length
|
| 463 |
+
if clip_seconds!=0:
|
| 464 |
+
print(f'###=====segment clip start, {round(len(dat) / audio_sr, 3)}s======')
|
| 465 |
+
# padd
|
| 466 |
+
pad_len = int(audio_sr * pad_seconds)
|
| 467 |
+
dat = np.concatenate([np.zeros([pad_len]), dat, np.zeros([pad_len])])
|
| 468 |
+
raw_path = io.BytesIO()
|
| 469 |
+
soundfile.write(raw_path, dat, audio_sr, format="wav")
|
| 470 |
+
raw_path.seek(0)
|
| 471 |
+
out_audio, out_sr, out_frame = self.infer(spk, tran, raw_path,
|
| 472 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
| 473 |
+
auto_predict_f0=auto_predict_f0,
|
| 474 |
+
noice_scale=noice_scale,
|
| 475 |
+
f0_predictor = f0_predictor,
|
| 476 |
+
enhancer_adaptive_key = enhancer_adaptive_key,
|
| 477 |
+
cr_threshold = cr_threshold,
|
| 478 |
+
k_step = k_step,
|
| 479 |
+
frame = global_frame,
|
| 480 |
+
spk_mix = use_spk_mix,
|
| 481 |
+
second_encoding = second_encoding,
|
| 482 |
+
loudness_envelope_adjustment = loudness_envelope_adjustment
|
| 483 |
+
)
|
| 484 |
+
global_frame += out_frame
|
| 485 |
+
_audio = out_audio.cpu().numpy()
|
| 486 |
+
pad_len = int(self.target_sample * pad_seconds)
|
| 487 |
+
_audio = _audio[pad_len:-pad_len]
|
| 488 |
+
_audio = pad_array(_audio, per_length)
|
| 489 |
+
if lg_size!=0 and k!=0:
|
| 490 |
+
lg1 = audio[-(lg_size_r+lg_size_c_r):-lg_size_c_r] if lgr_num != 1 else audio[-lg_size:]
|
| 491 |
+
lg2 = _audio[lg_size_c_l:lg_size_c_l+lg_size_r] if lgr_num != 1 else _audio[0:lg_size]
|
| 492 |
+
lg_pre = lg1*(1-lg)+lg2*lg
|
| 493 |
+
audio = audio[0:-(lg_size_r+lg_size_c_r)] if lgr_num != 1 else audio[0:-lg_size]
|
| 494 |
+
audio.extend(lg_pre)
|
| 495 |
+
_audio = _audio[lg_size_c_l+lg_size_r:] if lgr_num != 1 else _audio[lg_size:]
|
| 496 |
+
audio.extend(list(_audio))
|
| 497 |
+
return np.array(audio)
|
| 498 |
+
|
| 499 |
+
class RealTimeVC:
|
| 500 |
+
def __init__(self):
|
| 501 |
+
self.last_chunk = None
|
| 502 |
+
self.last_o = None
|
| 503 |
+
self.chunk_len = 16000 # chunk length
|
| 504 |
+
self.pre_len = 3840 # cross fade length, multiples of 640
|
| 505 |
+
|
| 506 |
+
# Input and output are 1-dimensional numpy waveform arrays
|
| 507 |
+
|
| 508 |
+
def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path,
|
| 509 |
+
cluster_infer_ratio=0,
|
| 510 |
+
auto_predict_f0=False,
|
| 511 |
+
noice_scale=0.4,
|
| 512 |
+
f0_filter=False):
|
| 513 |
+
|
| 514 |
+
import maad
|
| 515 |
+
audio, sr = torchaudio.load(input_wav_path)
|
| 516 |
+
audio = audio.cpu().numpy()[0]
|
| 517 |
+
temp_wav = io.BytesIO()
|
| 518 |
+
if self.last_chunk is None:
|
| 519 |
+
input_wav_path.seek(0)
|
| 520 |
+
|
| 521 |
+
audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path,
|
| 522 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
| 523 |
+
auto_predict_f0=auto_predict_f0,
|
| 524 |
+
noice_scale=noice_scale,
|
| 525 |
+
f0_filter=f0_filter)
|
| 526 |
+
|
| 527 |
+
audio = audio.cpu().numpy()
|
| 528 |
+
self.last_chunk = audio[-self.pre_len:]
|
| 529 |
+
self.last_o = audio
|
| 530 |
+
return audio[-self.chunk_len:]
|
| 531 |
+
else:
|
| 532 |
+
audio = np.concatenate([self.last_chunk, audio])
|
| 533 |
+
soundfile.write(temp_wav, audio, sr, format="wav")
|
| 534 |
+
temp_wav.seek(0)
|
| 535 |
+
|
| 536 |
+
audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav,
|
| 537 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
| 538 |
+
auto_predict_f0=auto_predict_f0,
|
| 539 |
+
noice_scale=noice_scale,
|
| 540 |
+
f0_filter=f0_filter)
|
| 541 |
+
|
| 542 |
+
audio = audio.cpu().numpy()
|
| 543 |
+
ret = maad.util.crossfade(self.last_o, audio, self.pre_len)
|
| 544 |
+
self.last_chunk = audio[-self.pre_len:]
|
| 545 |
+
self.last_o = audio
|
| 546 |
+
return ret[self.chunk_len:2 * self.chunk_len]
|
| 547 |
+
|
inference/slicer.py
ADDED
|
@@ -0,0 +1,142 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import librosa
|
| 2 |
+
import torch
|
| 3 |
+
import torchaudio
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
class Slicer:
|
| 7 |
+
def __init__(self,
|
| 8 |
+
sr: int,
|
| 9 |
+
threshold: float = -40.,
|
| 10 |
+
min_length: int = 5000,
|
| 11 |
+
min_interval: int = 300,
|
| 12 |
+
hop_size: int = 20,
|
| 13 |
+
max_sil_kept: int = 5000):
|
| 14 |
+
if not min_length >= min_interval >= hop_size:
|
| 15 |
+
raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size')
|
| 16 |
+
if not max_sil_kept >= hop_size:
|
| 17 |
+
raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size')
|
| 18 |
+
min_interval = sr * min_interval / 1000
|
| 19 |
+
self.threshold = 10 ** (threshold / 20.)
|
| 20 |
+
self.hop_size = round(sr * hop_size / 1000)
|
| 21 |
+
self.win_size = min(round(min_interval), 4 * self.hop_size)
|
| 22 |
+
self.min_length = round(sr * min_length / 1000 / self.hop_size)
|
| 23 |
+
self.min_interval = round(min_interval / self.hop_size)
|
| 24 |
+
self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
|
| 25 |
+
|
| 26 |
+
def _apply_slice(self, waveform, begin, end):
|
| 27 |
+
if len(waveform.shape) > 1:
|
| 28 |
+
return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)]
|
| 29 |
+
else:
|
| 30 |
+
return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)]
|
| 31 |
+
|
| 32 |
+
# @timeit
|
| 33 |
+
def slice(self, waveform):
|
| 34 |
+
if len(waveform.shape) > 1:
|
| 35 |
+
samples = librosa.to_mono(waveform)
|
| 36 |
+
else:
|
| 37 |
+
samples = waveform
|
| 38 |
+
if samples.shape[0] <= self.min_length:
|
| 39 |
+
return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
|
| 40 |
+
rms_list = librosa.feature.rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
|
| 41 |
+
sil_tags = []
|
| 42 |
+
silence_start = None
|
| 43 |
+
clip_start = 0
|
| 44 |
+
for i, rms in enumerate(rms_list):
|
| 45 |
+
# Keep looping while frame is silent.
|
| 46 |
+
if rms < self.threshold:
|
| 47 |
+
# Record start of silent frames.
|
| 48 |
+
if silence_start is None:
|
| 49 |
+
silence_start = i
|
| 50 |
+
continue
|
| 51 |
+
# Keep looping while frame is not silent and silence start has not been recorded.
|
| 52 |
+
if silence_start is None:
|
| 53 |
+
continue
|
| 54 |
+
# Clear recorded silence start if interval is not enough or clip is too short
|
| 55 |
+
is_leading_silence = silence_start == 0 and i > self.max_sil_kept
|
| 56 |
+
need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length
|
| 57 |
+
if not is_leading_silence and not need_slice_middle:
|
| 58 |
+
silence_start = None
|
| 59 |
+
continue
|
| 60 |
+
# Need slicing. Record the range of silent frames to be removed.
|
| 61 |
+
if i - silence_start <= self.max_sil_kept:
|
| 62 |
+
pos = rms_list[silence_start: i + 1].argmin() + silence_start
|
| 63 |
+
if silence_start == 0:
|
| 64 |
+
sil_tags.append((0, pos))
|
| 65 |
+
else:
|
| 66 |
+
sil_tags.append((pos, pos))
|
| 67 |
+
clip_start = pos
|
| 68 |
+
elif i - silence_start <= self.max_sil_kept * 2:
|
| 69 |
+
pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin()
|
| 70 |
+
pos += i - self.max_sil_kept
|
| 71 |
+
pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
|
| 72 |
+
pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
|
| 73 |
+
if silence_start == 0:
|
| 74 |
+
sil_tags.append((0, pos_r))
|
| 75 |
+
clip_start = pos_r
|
| 76 |
+
else:
|
| 77 |
+
sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
|
| 78 |
+
clip_start = max(pos_r, pos)
|
| 79 |
+
else:
|
| 80 |
+
pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
|
| 81 |
+
pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
|
| 82 |
+
if silence_start == 0:
|
| 83 |
+
sil_tags.append((0, pos_r))
|
| 84 |
+
else:
|
| 85 |
+
sil_tags.append((pos_l, pos_r))
|
| 86 |
+
clip_start = pos_r
|
| 87 |
+
silence_start = None
|
| 88 |
+
# Deal with trailing silence.
|
| 89 |
+
total_frames = rms_list.shape[0]
|
| 90 |
+
if silence_start is not None and total_frames - silence_start >= self.min_interval:
|
| 91 |
+
silence_end = min(total_frames, silence_start + self.max_sil_kept)
|
| 92 |
+
pos = rms_list[silence_start: silence_end + 1].argmin() + silence_start
|
| 93 |
+
sil_tags.append((pos, total_frames + 1))
|
| 94 |
+
# Apply and return slices.
|
| 95 |
+
if len(sil_tags) == 0:
|
| 96 |
+
return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
|
| 97 |
+
else:
|
| 98 |
+
chunks = []
|
| 99 |
+
# 第一段静音并非从头开始，补上有声片段
|
| 100 |
+
if sil_tags[0][0]:
|
| 101 |
+
chunks.append(
|
| 102 |
+
{"slice": False, "split_time": f"0,{min(waveform.shape[0], sil_tags[0][0] * self.hop_size)}"})
|
| 103 |
+
for i in range(0, len(sil_tags)):
|
| 104 |
+
# 标识有声片段（跳过第一段）
|
| 105 |
+
if i:
|
| 106 |
+
chunks.append({"slice": False,
|
| 107 |
+
"split_time": f"{sil_tags[i - 1][1] * self.hop_size},{min(waveform.shape[0], sil_tags[i][0] * self.hop_size)}"})
|
| 108 |
+
# 标识所有静音片段
|
| 109 |
+
chunks.append({"slice": True,
|
| 110 |
+
"split_time": f"{sil_tags[i][0] * self.hop_size},{min(waveform.shape[0], sil_tags[i][1] * self.hop_size)}"})
|
| 111 |
+
# 最后一段静音并非结尾，补上结尾片段
|
| 112 |
+
if sil_tags[-1][1] * self.hop_size < len(waveform):
|
| 113 |
+
chunks.append({"slice": False, "split_time": f"{sil_tags[-1][1] * self.hop_size},{len(waveform)}"})
|
| 114 |
+
chunk_dict = {}
|
| 115 |
+
for i in range(len(chunks)):
|
| 116 |
+
chunk_dict[str(i)] = chunks[i]
|
| 117 |
+
return chunk_dict
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
def cut(audio_path, db_thresh=-30, min_len=5000):
|
| 121 |
+
audio, sr = librosa.load(audio_path, sr=None)
|
| 122 |
+
slicer = Slicer(
|
| 123 |
+
sr=sr,
|
| 124 |
+
threshold=db_thresh,
|
| 125 |
+
min_length=min_len
|
| 126 |
+
)
|
| 127 |
+
chunks = slicer.slice(audio)
|
| 128 |
+
return chunks
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
def chunks2audio(audio_path, chunks):
|
| 132 |
+
chunks = dict(chunks)
|
| 133 |
+
audio, sr = torchaudio.load(audio_path)
|
| 134 |
+
if len(audio.shape) == 2 and audio.shape[1] >= 2:
|
| 135 |
+
audio = torch.mean(audio, dim=0).unsqueeze(0)
|
| 136 |
+
audio = audio.cpu().numpy()[0]
|
| 137 |
+
result = []
|
| 138 |
+
for k, v in chunks.items():
|
| 139 |
+
tag = v["split_time"].split(",")
|
| 140 |
+
if tag[0] != tag[1]:
|
| 141 |
+
result.append((v["slice"], audio[int(tag[0]):int(tag[1])]))
|
| 142 |
+
return result, sr
|