irebai commited on
Commit
a9bcfca
1 Parent(s): dd03b28

first commit

Browse files
Files changed (8) hide show
  1. README.md +123 -0
  2. config.json +76 -0
  3. preprocessor_config.json +8 -0
  4. pytorch_model.bin +3 -0
  5. refs.txt +0 -0
  6. special_tokens_map.json +1 -0
  7. trs.txt +0 -0
  8. vocab.json +1 -0
README.md ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: fr
3
+ datasets:
4
+ - common_voice
5
+ tags:
6
+ - audio
7
+ - automatic-speech-recognition
8
+ - speech
9
+ - xlsr-fine-tuning
10
+ license: apache-2.0
11
+ model-index:
12
+ - name: wav2vec2-large-xlsr-53-French_punctuation by Ilyes Rebai
13
+ results:
14
+ - task:
15
+ name: Speech Recognition
16
+ type: automatic-speech-recognition
17
+ dataset:
18
+ name: Common Voice
19
+ args: fr
20
+ metrics:
21
+ - name: Test WER and CER on text and puctuation prediction
22
+ types: [wer, cer]
23
+ values: [19.47%, 6.66%]
24
+ -name: Test WER and CER on text without punctuation
25
+ types: [wer, cer]
26
+ values: [17.88%, 6.37%]
27
+
28
+ ---
29
+ ## Evaluation on Common Voice FR Test
30
+ ```python
31
+ import re
32
+ import torch
33
+ import torchaudio
34
+ from datasets import load_dataset, load_metric
35
+ from transformers import (
36
+ Wav2Vec2ForCTC,
37
+ Wav2Vec2Processor,
38
+ )
39
+
40
+
41
+
42
+ model_name = "Ilyes/wav2vec2-large-xlsr-53-french_punctuation"
43
+
44
+
45
+ model = Wav2Vec2ForCTC.from_pretrained(model_name).to('cuda')
46
+ processor = Wav2Vec2Processor.from_pretrained(model_name)
47
+
48
+
49
+ ds = load_dataset("common_voice", "fr", split="test")
50
+
51
+
52
+ chars_to_ignore_regex = '[\;\:\"\“\%\‘\”\�\‘\’\’\’\‘\…\·\ǃ\«\‹\»\›“\”\\ʿ\ʾ\„\∞\\|\;\:\*\—\–\─\―\_\/\:\ː\;\=\«\»\→]'
53
+ def normalize_text(text):
54
+ text = text.lower().strip()
55
+ text = re.sub('œ', 'oe', text)
56
+ text = re.sub('æ', 'ae', text)
57
+ text = re.sub("’|´|′|ʼ|‘|ʻ|`", "'", text)
58
+ text = re.sub("'+ ", " ", text)
59
+ text = re.sub(" '+", " ", text)
60
+ text = re.sub("'$", " ", text)
61
+ text = re.sub("' ", " ", text)
62
+ text = re.sub("−|‐", "-", text)
63
+ text = re.sub(" -", "", text)
64
+ text = re.sub("- ", "", text)
65
+ text = re.sub(chars_to_ignore_regex, '', text)
66
+ return text
67
+
68
+
69
+
70
+ def map_to_array(batch):
71
+ speech, _ = torchaudio.load(batch["path"])
72
+ batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
73
+ batch["sampling_rate"] = resampler.new_freq
74
+ batch["sentence"] = normalize_text(batch["sentence"])
75
+ return batch
76
+
77
+ ds = ds.map(map_to_array)
78
+
79
+ resampler = torchaudio.transforms.Resample(48_000, 16_000)
80
+ def map_to_pred(batch):
81
+ features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
82
+ input_values = features.input_values.to(device)
83
+ attention_mask = features.attention_mask.to(device)
84
+ with torch.no_grad():
85
+ logits = model(input_values, attention_mask=attention_mask).logits
86
+ pred_ids = torch.argmax(logits, dim=-1)
87
+ batch["predicted"] = processor.batch_decode(pred_ids)
88
+ batch["target"] = batch["sentence"]
89
+ # remove duplicates
90
+ batch["target"] = re.sub('\.+', '.', batch["target"])
91
+ batch["target"] = re.sub('\?+', '?', batch["target"])
92
+ batch["target"] = re.sub('!+', '!', batch["target"])
93
+ batch["target"] = re.sub(',+', ',', batch["target"])
94
+ return batch
95
+
96
+ result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
97
+ wer = load_metric("wer")
98
+ print(wer.compute(predictions=result["predicted"], references=result["target"]))
99
+ ```
100
+ ## Some results
101
+
102
+ | Reference | Prediction |
103
+ | ------------- | ------------- |
104
+ | il vécut à new york et y enseigna une grande partie de sa vie. | il a vécu à new york et y enseigna une grande partie de sa vie. |
105
+ | au classement par nations, l'allemagne est la tenante du titre. | au classement der nation l'allemagne est la tenante du titre.. |
106
+ | voici un petit calcul pour fixer les idées. | voici un petit calcul pour fixer les idées. |
107
+ | oh! tu dois être beau avec | oh! tu dois être beau avec. |
108
+ | babochet vous le voulez? | baboche, vous le voulez? |
109
+ | la commission est, par conséquent, défavorable à cet amendement. | la commission est, par conséquent, défavorable à cet amendement. |
110
+
111
+ All the references and predictions of the test corpus are already available in this repository.
112
+
113
+ ## Results
114
+
115
+ text + punctuation
116
+
117
+ WER=21.47% CER=7.21%
118
+
119
+
120
+ text (without punctuation)
121
+
122
+ WER=19.71% CER=6.91%
123
+
config.json ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/workspace/output_models/fr/wav2vec2-large-xlsr-53/checkpoint-14700",
3
+ "activation_dropout": 0.055,
4
+ "apply_spec_augment": true,
5
+ "architectures": [
6
+ "Wav2Vec2ForCTC"
7
+ ],
8
+ "attention_dropout": 0.094,
9
+ "bos_token_id": 1,
10
+ "conv_bias": true,
11
+ "conv_dim": [
12
+ 512,
13
+ 512,
14
+ 512,
15
+ 512,
16
+ 512,
17
+ 512,
18
+ 512
19
+ ],
20
+ "conv_kernel": [
21
+ 10,
22
+ 3,
23
+ 3,
24
+ 3,
25
+ 3,
26
+ 2,
27
+ 2
28
+ ],
29
+ "conv_stride": [
30
+ 5,
31
+ 2,
32
+ 2,
33
+ 2,
34
+ 2,
35
+ 2,
36
+ 2
37
+ ],
38
+ "ctc_loss_reduction": "mean",
39
+ "ctc_zero_infinity": false,
40
+ "do_stable_layer_norm": true,
41
+ "eos_token_id": 2,
42
+ "feat_extract_activation": "gelu",
43
+ "feat_extract_dropout": 0.0,
44
+ "feat_extract_norm": "layer",
45
+ "feat_proj_dropout": 0.04,
46
+ "final_dropout": 0.0,
47
+ "gradient_checkpointing": true,
48
+ "hidden_act": "gelu",
49
+ "hidden_dropout": 0.047,
50
+ "hidden_size": 1024,
51
+ "initializer_range": 0.02,
52
+ "intermediate_size": 4096,
53
+ "layer_norm_eps": 1e-05,
54
+ "layerdrop": 0.041,
55
+ "mask_channel_length": 10,
56
+ "mask_channel_min_space": 1,
57
+ "mask_channel_other": 0.0,
58
+ "mask_channel_prob": 0.0,
59
+ "mask_channel_selection": "static",
60
+ "mask_feature_length": 10,
61
+ "mask_feature_prob": 0.0,
62
+ "mask_time_length": 10,
63
+ "mask_time_min_space": 1,
64
+ "mask_time_other": 0.0,
65
+ "mask_time_prob": 0.4,
66
+ "mask_time_selection": "static",
67
+ "model_type": "wav2vec2",
68
+ "num_attention_heads": 16,
69
+ "num_conv_pos_embedding_groups": 16,
70
+ "num_conv_pos_embeddings": 128,
71
+ "num_feat_extract_layers": 7,
72
+ "num_hidden_layers": 24,
73
+ "pad_token_id": 48,
74
+ "transformers_version": "4.5.0.dev0",
75
+ "vocab_size": 49
76
+ }
preprocessor_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": true,
3
+ "feature_size": 1,
4
+ "padding_side": "right",
5
+ "padding_value": 0.0,
6
+ "return_attention_mask": true,
7
+ "sampling_rate": 16000
8
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9f0c763c2c455ca6fa21a4644e7e37991dee8d5e436d5dcb73e21d9acd932275
3
+ size 1262134743
refs.txt ADDED
The diff for this file is too large to render. See raw diff
 
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}
trs.txt ADDED
The diff for this file is too large to render. See raw diff
 
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"a": 0, "e": 1, "i": 2, "o": 3, "u": 4, "y": 5, "b": 6, "c": 7, "d": 8, "f": 9, "g": 10, "h": 11, "j": 12, "k": 13, "l": 14, "m": 15, "n": 16, "p": 17, "q": 18, "r": 19, "s": 20, "t": 21, "v": 22, "w": 23, "x": 24, "z": 25, "à": 26, "â": 27, "ç": 28, "è": 29, "é": 30, "ê": 31, "ë": 32, "î": 33, "ï": 34, "ô": 35, "ù": 36, "û": 37, "ü": 38, "ÿ": 39, "|": 40, "'": 41, "-": 42, ".": 43, ",": 44, "!": 45, "?": 46, "<unk>": 47, "<pad>": 48}