jonatasgrosman commited on
Commit
ec0a4d2
1 Parent(s): ac5e62b

update model

Browse files

Files changed (4) hide show
  1. README.md +15 -80
  2. config.json +1 -1
  3. preprocessor_config.json +1 -0
  4. pytorch_model.bin +1 -1
README.md CHANGED
@@ -2,8 +2,6 @@
2
  language: en
3
  datasets:
4
  - common_voice
5
- - librispeech_asr
6
- - timit_asr
7
  metrics:
8
  - wer
9
  - cer
@@ -26,15 +24,15 @@ model-index:
26
  metrics:
27
  - name: Test WER
28
  type: wer
29
- value: 19.76
30
  - name: Test CER
31
  type: cer
32
- value: 8.60
33
  ---
34
 
35
  # Wav2Vec2-Large-XLSR-53-English
36
 
37
- 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), [LibriSpeech](https://huggingface.co/datasets/librispeech_asr) and [TIMIT](https://huggingface.co/datasets/timit_asr),.
38
  When using this model, make sure that your speech input is sampled at 16kHz.
39
 
40
  The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint
@@ -83,16 +81,16 @@ for i, predicted_sentence in enumerate(predicted_sentences):
83
 
84
  | Reference | Prediction |
85
  | ------------- | ------------- |
86
- | "SHE'LL BE ALL RIGHT." | SHE'D BE ALRIGHT |
87
  | SIX | SIX |
88
- | "ALL'S WELL THAT ENDS WELL." | ALL IS WELL THAT ENDS WELL |
89
  | DO YOU MEAN IT? | DO YOU MEAN IT |
90
  | 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 |
91
- | HOW IS MOZILLA GOING TO HANDLE AMBIGUITIES LIKE QUEUE AND CUE? | HOW IS MUSILA GOING TO HANDLE ANB HOOTIES LIKE QU AND QU |
92
- | "I GUESS YOU MUST THINK I'M KINDA BATTY." | RISIONAS INCI IN TE BACTY |
93
  | NO ONE NEAR THE REMOTE MACHINE YOU COULD RING? | NO ONE NEAR THE REMOTE MACHINE YOU COULD RING |
94
- | SAUCE FOR THE GOOSE IS SAUCE FOR THE GANDER. | SAUCE FOR THE GUISE IS SAUCE FOR THE GONDER |
95
- | GROVES STARTED WRITING SONGS WHEN SHE WAS FOUR YEARS OLD. | GRAFS STARTED WRITING SOUNDS WHEN SHE WAS FOUR YEARS OLD |
96
 
97
  ## Evaluation
98
 
@@ -117,11 +115,6 @@ CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '
117
 
118
  test_dataset = load_dataset("common_voice", LANG_ID, split="test")
119
 
120
- # uncomment the following lines to eval using other datasets
121
- # test_dataset = load_dataset("librispeech_asr", "clean", split="test")
122
- # test_dataset = load_dataset("librispeech_asr", "other", split="test")
123
- # test_dataset = load_dataset("timit_asr", split="test")
124
-
125
  wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
126
  cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py
127
 
@@ -136,9 +129,9 @@ model.to(DEVICE)
136
  def speech_file_to_array_fn(batch):
137
  with warnings.catch_warnings():
138
  warnings.simplefilter("ignore")
139
- speech_array, sampling_rate = librosa.load(batch["file"] if "file" in batch else batch["path"], sr=16_000)
140
  batch["speech"] = speech_array
141
- batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["text"] if "text" in batch else batch["sentence"]).upper()
142
  return batch
143
 
144
  test_dataset = test_dataset.map(speech_file_to_array_fn)
@@ -166,76 +159,18 @@ print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_
166
 
167
  **Test Result**:
168
 
169
- 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. Initially, I've tested the model only using the Common Voice dataset. Later I've also tested the model using the LibriSpeech and TIMIT datasets, which are better-behaved datasets than the Common Voice, containing only examples in US English extracted from audiobooks.
170
-
171
- ---
172
-
173
- **Common Voice**
174
 
175
  | Model | WER | CER |
176
  | ------------- | ------------- | ------------- |
177
- | jonatasgrosman/wav2vec2-large-xlsr-53-english | **19.76%** | **8.60%** |
178
- | jonatasgrosman/wav2vec2-large-english | 21.16% | 9.53% |
179
  | facebook/wav2vec2-large-960h-lv60-self | 22.03% | 10.39% |
180
  | facebook/wav2vec2-large-960h-lv60 | 23.97% | 11.14% |
 
181
  | facebook/wav2vec2-large-960h | 32.79% | 16.03% |
182
- | boris/xlsr-en-punctuation | 34.81% | 15.51% |
183
  | facebook/wav2vec2-base-960h | 39.86% | 19.89% |
184
  | facebook/wav2vec2-base-100h | 51.06% | 25.06% |
185
  | elgeish/wav2vec2-large-lv60-timit-asr | 59.96% | 34.28% |
186
  | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 66.41% | 36.76% |
187
  | elgeish/wav2vec2-base-timit-asr | 68.78% | 36.81% |
188
-
189
- ---
190
-
191
- **LibriSpeech (clean)**
192
-
193
- | Model | WER | CER |
194
- | ------------- | ------------- | ------------- |
195
- | facebook/wav2vec2-large-960h-lv60-self | **1.86%** | **0.54%** |
196
- | facebook/wav2vec2-large-960h-lv60 | 2.15% | 0.61% |
197
- | facebook/wav2vec2-large-960h | 2.82% | 0.84% |
198
- | facebook/wav2vec2-base-960h | 3.44% | 1.06% |
199
- | jonatasgrosman/wav2vec2-large-xlsr-53-english | 4.16% | 1.28% |
200
- | facebook/wav2vec2-base-100h | 6.26% | 2.00% |
201
- | jonatasgrosman/wav2vec2-large-english | 8.00% | 2.55% |
202
- | elgeish/wav2vec2-large-lv60-timit-asr | 15.53% | 4.93% |
203
- | boris/xlsr-en-punctuation | 19.28% | 6.45% |
204
- | elgeish/wav2vec2-base-timit-asr | 29.19% | 8.38% |
205
- | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 31.82% | 12.41% |
206
-
207
- ---
208
-
209
- **LibriSpeech (other)**
210
-
211
- | Model | WER | CER |
212
- | ------------- | ------------- | ------------- |
213
- | facebook/wav2vec2-large-960h-lv60-self | **3.89%** | **1.40%** |
214
- | facebook/wav2vec2-large-960h-lv60 | 4.45% | 1.56% |
215
- | facebook/wav2vec2-large-960h | 6.49% | 2.52% |
216
- | jonatasgrosman/wav2vec2-large-xlsr-53-english | 8.82% | 3.42% |
217
- | facebook/wav2vec2-base-960h | 8.90% | 3.55% |
218
- | jonatasgrosman/wav2vec2-large-english | 13.62% | 5.24% |
219
- | facebook/wav2vec2-base-100h | 13.97% | 5.51% |
220
- | boris/xlsr-en-punctuation | 26.40% | 10.11% |
221
- | elgeish/wav2vec2-large-lv60-timit-asr | 28.39% | 12.08% |
222
- | elgeish/wav2vec2-base-timit-asr | 42.04% | 15.57% |
223
- | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 45.19% | 20.32% |
224
-
225
- ---
226
-
227
- **TIMIT**
228
-
229
- | Model | WER | CER |
230
- | ------------- | ------------- | ------------- |
231
- | facebook/wav2vec2-large-960h-lv60-self | **5.17%** | **1.33%** |
232
- | facebook/wav2vec2-large-960h-lv60 | 6.24% | 1.54% |
233
- | jonatasgrosman/wav2vec2-large-xlsr-53-english | 6.81% | 2.02% |
234
- | facebook/wav2vec2-large-960h | 9.63% | 2.19% |
235
- | facebook/wav2vec2-base-960h | 11.48% | 2.76% |
236
- | elgeish/wav2vec2-large-lv60-timit-asr | 13.83% | 4.36% |
237
- | jonatasgrosman/wav2vec2-large-english | 13.91% | 4.01% |
238
- | facebook/wav2vec2-base-100h | 16.75% | 4.79% |
239
- | elgeish/wav2vec2-base-timit-asr | 25.40% | 8.16% |
240
- | boris/xlsr-en-punctuation | 25.93% | 9.99% |
241
- | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 51.08% | 19.84% |
2
  language: en
3
  datasets:
4
  - common_voice
 
 
5
  metrics:
6
  - wer
7
  - cer
24
  metrics:
25
  - name: Test WER
26
  type: wer
27
+ value: 18.98
28
  - name: Test CER
29
  type: cer
30
+ value: 8.29
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
81
 
82
  | Reference | Prediction |
83
  | ------------- | ------------- |
84
+ | "SHE'LL BE ALL RIGHT." | SHE'LL BE ALL RIGHT |
85
  | SIX | SIX |
86
+ | "ALL'S WELL THAT ENDS WELL." | ALL AS 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? | HOW IS MOSLILLAR GOING TO HANDLE ANDBEWOOTH HIS LIKE Q AND Q |
90
+ | "I GUESS YOU MUST THINK I'M KINDA BATTY." | RUSTIAN WASTIN PAN ONTE BATTLY |
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 GUICE IS SAUCE 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
 
115
 
116
  test_dataset = load_dataset("common_voice", LANG_ID, split="test")
117
 
 
 
 
 
 
118
  wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
119
  cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py
120
 
129
  def speech_file_to_array_fn(batch):
130
  with warnings.catch_warnings():
131
  warnings.simplefilter("ignore")
132
+ speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
133
  batch["speech"] = speech_array
134
+ batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper()
135
  return batch
136
 
137
  test_dataset = test_dataset.map(speech_file_to_array_fn)
159
 
160
  **Test Result**:
161
 
162
+ 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 2021-06-17). 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.
 
 
 
 
163
 
164
  | Model | WER | CER |
165
  | ------------- | ------------- | ------------- |
166
+ | jonatasgrosman/wav2vec2-large-xlsr-53-english | **18.98%** | **8.29%** |
167
+ | jonatasgrosman/wav2vec2-large-english | 21.53% | 9.66% |
168
  | facebook/wav2vec2-large-960h-lv60-self | 22.03% | 10.39% |
169
  | facebook/wav2vec2-large-960h-lv60 | 23.97% | 11.14% |
170
+ | boris/xlsr-en-punctuation | 29.10% | 10.75% |
171
  | facebook/wav2vec2-large-960h | 32.79% | 16.03% |
 
172
  | facebook/wav2vec2-base-960h | 39.86% | 19.89% |
173
  | facebook/wav2vec2-base-100h | 51.06% | 25.06% |
174
  | elgeish/wav2vec2-large-lv60-timit-asr | 59.96% | 34.28% |
175
  | facebook/wav2vec2-base-10k-voxpopuli-ft-en | 66.41% | 36.76% |
176
  | elgeish/wav2vec2-base-timit-asr | 68.78% | 36.81% |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
config.json CHANGED
@@ -71,6 +71,6 @@
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
  }
71
  "num_feat_extract_layers": 7,
72
  "num_hidden_layers": 24,
73
  "pad_token_id": 0,
74
+ "transformers_version": "4.7.0.dev0",
75
  "vocab_size": 33
76
  }
preprocessor_config.json CHANGED
@@ -1,5 +1,6 @@
1
  {
2
  "do_normalize": true,
 
3
  "feature_size": 1,
4
  "padding_side": "right",
5
  "padding_value": 0.0,
1
  {
2
  "do_normalize": true,
3
+ "feature_extractor_type": "Wav2Vec2FeatureExtractor",
4
  "feature_size": 1,
5
  "padding_side": "right",
6
  "padding_value": 0.0,
pytorch_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f9a3cc8402adf407944f67025114dd0f02af9584bfb266c031f586d601719b0a
3
  size 1262069143
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b7688644eeefe1f5760bb4c4a61d085793a3740159fdbf19fd37c5d4f3729bf
3
  size 1262069143