jonatasgrosman commited on
Commit
64967ea
1 Parent(s): c8c756d

first commit

Browse files
README.md ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ datasets:
4
+ - common_voice
5
+ metrics:
6
+ - wer
7
+ - cer
8
+ tags:
9
+ - audio
10
+ - automatic-speech-recognition
11
+ - speech
12
+ - xlsr-fine-tuning-week
13
+ license: apache-2.0
14
+ model-index:
15
+ - name: XLSR Wav2Vec2 English by Jonatas Grosman
16
+ results:
17
+ - task:
18
+ name: Speech Recognition
19
+ type: automatic-speech-recognition
20
+ dataset:
21
+ name: Common Voice en
22
+ type: common_voice
23
+ args: en
24
+ metrics:
25
+ - name: Test WER
26
+ type: wer
27
+ value: 39.59
28
+ - name: Test CER
29
+ type: cer
30
+ value: 18.18
31
+ ---
32
+
33
+ # Wav2Vec2-Large-XLSR-53-English
34
+
35
+ Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on English using the [Common Voice](https://huggingface.co/datasets/common_voice).
36
+ When using this model, make sure that your speech input is sampled at 16kHz.
37
+
38
+ The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint
39
+
40
+ ## Usage
41
+
42
+ The model can be used directly (without a language model) as follows:
43
+
44
+ ```python
45
+ import torch
46
+ import librosa
47
+ from datasets import load_dataset
48
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
49
+
50
+ LANG_ID = "en"
51
+ MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
52
+ SAMPLES = 10
53
+
54
+ test_dataset = load_dataset("common_voice", LANG_ID, split=f"test[:{SAMPLES}]")
55
+
56
+ processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
57
+ model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
58
+
59
+ # Preprocessing the datasets.
60
+ # We need to read the audio files as arrays
61
+ def speech_file_to_array_fn(batch):
62
+ speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
63
+ batch["speech"] = speech_array
64
+ batch["sentence"] = batch["sentence"].upper()
65
+ return batch
66
+
67
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
68
+ inputs = processor(test_dataset["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
69
+
70
+ with torch.no_grad():
71
+ logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
72
+
73
+ predicted_ids = torch.argmax(logits, dim=-1)
74
+ predicted_sentences = processor.batch_decode(predicted_ids)
75
+
76
+ for i, predicted_sentence in enumerate(predicted_sentences):
77
+ print("-" * 100)
78
+ print("Reference:", test_dataset[i]["sentence"])
79
+ print("Prediction:", predicted_sentence)
80
+ ```
81
+
82
+ | Reference | Prediction |
83
+ | ------------- | ------------- |
84
+ | "SHE'LL BE ALL RIGHT." | SHE'LD BE ALL RIGHT |
85
+ | SIX | SIX |
86
+ | "ALL'S WELL THAT ENDS WELL." | ALL IS WELL THAT ENDS WELL |
87
+ | DO YOU MEAN IT? | DO YOU MEAN IT |
88
+ | THE NEW PATCH IS LESS INVASIVE THAN THE OLD ONE, BUT STILL CAUSES REGRESSIONS. | THE NEW PATCH IS LESS INVASIVE THAN THE OLD ONE BUT STILL CAUSES REGRESSION |
89
+ | HOW IS MOZILLA GOING TO HANDLE AMBIGUITIES LIKE QUEUE AND CUE? | HOWIS MOCILE ARE GOING TO HANDLE AMBIGUITIES LIKE KU AND KU |
90
+ | "I GUESS YOU MUST THINK I'M KINDA BATTY." | RISSHON WAS INCAN IN THE BAK TE |
91
+ | NO ONE NEAR THE REMOTE MACHINE YOU COULD RING? | NO ONE NEAR THE REMOTE MACHINE YOU COULD RING |
92
+ | SAUCE FOR THE GOOSE IS SAUCE FOR THE GANDER. | SAUCE FOR THE GUISE IS SAUCED FOR THE GONDER |
93
+ | GROVES STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD. | GRAFS STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD |
94
+
95
+ ## Evaluation
96
+
97
+ The model can be evaluated as follows on the English test data of Common Voice.
98
+
99
+ ```python
100
+ import torch
101
+ import re
102
+ import librosa
103
+ from datasets import load_dataset, load_metric
104
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
105
+
106
+ LANG_ID = "en"
107
+ MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-english"
108
+ DEVICE = "cuda"
109
+
110
+ CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
111
+ "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
112
+ "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
113
+ "、", "﹂", "﹁", "‧", "~", "﹏", ",", "{", "}", "(", ")", "[", "]", "【", "】", "‥", "〽",
114
+ "『", "』", "〝", "〟", "⟨", "⟩", "〜", ":", "!", "?", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
115
+
116
+ test_dataset = load_dataset("common_voice", LANG_ID, split="test")
117
+
118
+ # uncomment the following lines to eval using other datasets
119
+ # test_dataset = load_dataset("librispeech_asr", "clean", split="test")
120
+ # test_dataset = load_dataset("librispeech_asr", "other", split="test")
121
+ # test_dataset = load_dataset("timit_asr", split="test")
122
+
123
+ wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
124
+ cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py
125
+
126
+ chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
127
+
128
+ processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
129
+ model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
130
+ model.to(DEVICE)
131
+
132
+ # Preprocessing the datasets.
133
+ # We need to read the audio files as arrays
134
+ def speech_file_to_array_fn(batch):
135
+ with warnings.catch_warnings():
136
+ warnings.simplefilter("ignore")
137
+ speech_array, sampling_rate = librosa.load(batch["file"] if "file" in batch else batch["path"], sr=16_000)
138
+ batch["speech"] = speech_array
139
+ batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["text"] if "text" in batch else batch["sentence"]).upper()
140
+ return batch
141
+
142
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
143
+
144
+ # Preprocessing the datasets.
145
+ # We need to read the audio files as arrays
146
+ def evaluate(batch):
147
+ inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
148
+
149
+ with torch.no_grad():
150
+ logits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits
151
+
152
+ pred_ids = torch.argmax(logits, dim=-1)
153
+ batch["pred_strings"] = processor.batch_decode(pred_ids)
154
+ return batch
155
+
156
+ result = test_dataset.map(evaluate, batched=True, batch_size=8)
157
+
158
+ predictions = [x.upper() for x in result["pred_strings"]]
159
+ references = [x.upper() for x in result["sentence"]]
160
+
161
+ print(f"WER: {wer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
162
+ print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
163
+ ```
164
+
165
+ **Test Result**:
166
+
167
+ In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the evaluation script described above on other models as well (on 2021-05-20). Note that the table below may show different results from those already reported, this may have been caused due to some specificity of the other evaluation scripts used.
168
+
169
+ ---
170
+
171
+ **Common Voice**
172
+
173
+ | Model | WER | CER |
174
+ | ------------- | ------------- | ------------- |
175
+ | jonatasgrosman/wav2vec2-large-xlsr-53-english | **19.18%** | **8.25%** |
176
+ | jonatasgrosman/wav2vec2-large-english | 21.16% | 9.53% |
177
+ | facebook/wav2vec2-large-960h-lv60-self | 22.03% | 10.39% |
178
+ | facebook/wav2vec2-large-960h-lv60 | 23.97% | 11.14% |
179
+ | facebook/wav2vec2-large-960h | 32.79% | 16.03% |
180
+ | boris/xlsr-en-punctuation | 34.81% | 15.51% |
181
+ | facebook/wav2vec2-base-960h | 39.86% | 19.89% |
182
+ | facebook/wav2vec2-base-100h | 51.06% | 25.06% |
183
+ | elgeish/wav2vec2-large-lv60-timit-asr | 59.96% | 34.28% |
184
+ | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 66.41% | 36.76% |
185
+ | elgeish/wav2vec2-base-timit-asr | 68.78% | 36.81% |
186
+
187
+ ---
188
+
189
+ **LibriSpeech (clean)**
190
+
191
+ | Model | WER | CER |
192
+ | ------------- | ------------- | ------------- |
193
+ | facebook/wav2vec2-large-960h-lv60-self | **1.86%** | **0.54%** |
194
+ | facebook/wav2vec2-large-960h-lv60 | 2.15% | 0.61% |
195
+ | facebook/wav2vec2-large-960h | 2.82% | 0.84% |
196
+ | facebook/wav2vec2-base-960h | 3.44% | 1.06% |
197
+ | facebook/wav2vec2-base-100h | 6.26% | 2.00% |
198
+ | jonatasgrosman/wav2vec2-large-xlsr-53-english | 6.97% | 2.02% |
199
+ | jonatasgrosman/wav2vec2-large-english | 8.00% | 2.55% |
200
+ | elgeish/wav2vec2-large-lv60-timit-asr | 15.53% | 4.93% |
201
+ | boris/xlsr-en-punctuation | 19.28% | 6.45% |
202
+ | elgeish/wav2vec2-base-timit-asr | 29.19% | 8.38% |
203
+ | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 31.82% | 12.41% |
204
+
205
+ ---
206
+
207
+ **LibriSpeech (other)**
208
+
209
+ | Model | WER | CER |
210
+ | ------------- | ------------- | ------------- |
211
+ | facebook/wav2vec2-large-960h-lv60-self | **3.89%** | **1.40%** |
212
+ | facebook/wav2vec2-large-960h-lv60 | 4.45% | 1.56% |
213
+ | facebook/wav2vec2-large-960h | 6.49% | 2.52% |
214
+ | facebook/wav2vec2-base-960h | 8.90% | 3.55% |
215
+ | jonatasgrosman/wav2vec2-large-xlsr-53-english | 11.75% | 4.23% |
216
+ | jonatasgrosman/wav2vec2-large-english | 13.62% | 5.24% |
217
+ | facebook/wav2vec2-base-100h | 13.97% | 5.51% |
218
+ | boris/xlsr-en-punctuation | 26.40% | 10.11% |
219
+ | elgeish/wav2vec2-large-lv60-timit-asr | 28.39% | 12.08% |
220
+ | elgeish/wav2vec2-base-timit-asr | 42.04% | 15.57% |
221
+ | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 45.19% | 20.32% |
222
+
223
+ ---
224
+
225
+ **TIMIT**
226
+
227
+ | Model | WER | CER |
228
+ | ------------- | ------------- | ------------- |
229
+ | facebook/wav2vec2-large-960h-lv60-self | **5.17%** | **1.33%** |
230
+ | facebook/wav2vec2-large-960h-lv60 | 6.24% | 1.54% |
231
+ | facebook/wav2vec2-large-960h | 9.63% | 2.19% |
232
+ | facebook/wav2vec2-base-960h | 11.48% | 2.76% |
233
+ | jonatasgrosman/wav2vec2-large-xlsr-53-english | 11.93% | 3.50% |
234
+ | elgeish/wav2vec2-large-lv60-timit-asr | 13.83% | 4.36% |
235
+ | jonatasgrosman/wav2vec2-large-english | 13.91% | 4.01% |
236
+ | facebook/wav2vec2-base-100h | 16.75% | 4.79% |
237
+ | elgeish/wav2vec2-base-timit-asr | 25.40% | 8.16% |
238
+ | boris/xlsr-en-punctuation | 25.93% | 9.99% |
239
+ | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 51.08% | 19.84% |
config.json ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
3
+ "activation_dropout": 0.05,
4
+ "apply_spec_augment": true,
5
+ "architectures": [
6
+ "Wav2Vec2ForCTC"
7
+ ],
8
+ "attention_dropout": 0.1,
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": true,
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": 0.05,
46
+ "final_dropout": 0.0,
47
+ "gradient_checkpointing": true,
48
+ "hidden_act": "gelu",
49
+ "hidden_dropout": 0.05,
50
+ "hidden_size": 1024,
51
+ "initializer_range": 0.02,
52
+ "intermediate_size": 4096,
53
+ "layer_norm_eps": 1e-05,
54
+ "layerdrop": 0.05,
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": 0.05,
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": 0,
74
+ "transformers_version": "4.5.0.dev0",
75
+ "vocab_size": 33
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 sha256:3ba6ad16f6ecfcadd07ca5bc3353c3168665bee9fbfb160fbc864a6a9f87ca58
3
+ size 1262069143
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<pad>": 0, "<s>": 1, "</s>": 2, "<unk>": 3, "|": 4, "'": 5, "-": 6, "A": 7, "B": 8, "C": 9, "D": 10, "E": 11, "F": 12, "G": 13, "H": 14, "I": 15, "J": 16, "K": 17, "L": 18, "M": 19, "N": 20, "O": 21, "P": 22, "Q": 23, "R": 24, "S": 25, "T": 26, "U": 27, "V": 28, "W": 29, "X": 30, "Y": 31, "Z": 32}