Leonoravesterbacka commited on
Commit
6822ca7
·
verified ·
1 Parent(s): a03c6d9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +46 -11
README.md CHANGED
@@ -216,40 +216,75 @@ Stage 1 employed low threshold values (0 to 0.30 BLEU depending on dataset), whe
216
  | NST | 250 | 250 |
217
  | **Total** | **56,514** | **8,533** |
218
 
219
- The default when loading our models through Hugging Face is **Stage 2**. We have however also uploaded continued pretraining checkpoints and tagged them. You can load these other checkpoints by specifying the `revision` in `.from_pretrained()`. The pretrained checkpoints tag can for example be found here: [`pretrained-checkpoint`](https://huggingface.co/KBLab/kb-whisper-small/tree/pretrained-checkpoint). The Stage 2 default model tag is named `standard`. We supply a different stage 2 checkpoint -- with a more condensed style of transcribing -- under the name `subtitle`.
220
 
221
  ### Evaluation
222
 
223
 
224
- #### WER
225
  | Model size | | FLEURS | CommonVoice | NST |
226
  |------------|---------|--------|-------------|------|
227
  | [tiny](https://huggingface.co/KBLab/kb-whisper-tiny) | **KBLab** | **13.2** | **12.9** | **11.2** |
228
  | | OpenAI | 59.2 | 67.8 | 85.2 |
229
  | [base](https://huggingface.co/KBLab/kb-whisper-base) | **KBLab** | **9.1** | **8.7** | **7.8** |
230
  | | OpenAI | 39.6 | 52.1 | 53.4 |
231
- | [small](https://huggingface.co/KBLab/kb-whisper-small) | **KBLab** | **7.3** | **6.4** | **6.6** |
232
  | | OpenAI | 20.6 | 26.4 | 26.4 |
233
- | [medium](https://huggingface.co/KBLab/kb-whisper-medium) | **KBLab** | **6.6** | **5.4** | **5.8** |
234
  | | OpenAI | 12.1 | 15.8 | 17.1 |
235
- | [large-v3](https://huggingface.co/KBLab/kb-whisper-large) | **KBLab** | **5.4** | **4.1** | **5.2** |
236
  | | OpenAI | 7.8 | 9.5 | 11.3 |
237
 
 
238
 
239
- #### BLEU Score
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
240
  | Model size | | FLEURS | CommonVoice | NST |
241
  |------------|---------|--------|-------------|------|
242
- | tiny | KBLab | **76.6** | **73.7** | **74.3** |
243
  | | OpenAI | 26.9 | 21.1 | 24.0 |
244
- | base | KBLab | **83.2** | **79.9** | **78.3** |
245
  | | OpenAI | 41.1 | 32.5 | 36.9 |
246
- | small | KBLab | **86.6** | **83.5** | **79.6** |
247
  | | OpenAI | 64.0 | 56.5 | 58.2 |
248
- | medium | KBLab | **87.6** | **85.0** | **80.2** |
249
  | | OpenAI | 77.1 | 70.1 | 68.9 |
250
- | large-v3 | KBLab | **89.8** | **87.2** | **81.1** |
251
  | | OpenAI | 84.9 | 79.1 | 75.1 |
252
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
253
 
254
  ### Acknowledgements
255
 
 
216
  | NST | 250 | 250 |
217
  | **Total** | **56,514** | **8,533** |
218
 
219
+ The default when loading our models through Hugging Face is **Stage 2**. We have however also uploaded continued pretraining checkpoints and tagged them. You can load these other checkpoints by specifying the `revision` in `.from_pretrained()`. The pretrained checkpoints tag can for example be found here: [`pretrained-checkpoint`](https://huggingface.co/KBLab/kb-whisper-large/tree/pretrained-checkpoint). The Stage 2 default model tag is named `standard`. We supply two different stage 2 checkpoints -- one with a more condensed style of transcribing -- under the name `subtitle`, and one more verbose called `strict`.
220
 
221
  ### Evaluation
222
 
223
 
224
+ #### WER compared to OpenAI
225
  | Model size | | FLEURS | CommonVoice | NST |
226
  |------------|---------|--------|-------------|------|
227
  | [tiny](https://huggingface.co/KBLab/kb-whisper-tiny) | **KBLab** | **13.2** | **12.9** | **11.2** |
228
  | | OpenAI | 59.2 | 67.8 | 85.2 |
229
  | [base](https://huggingface.co/KBLab/kb-whisper-base) | **KBLab** | **9.1** | **8.7** | **7.8** |
230
  | | OpenAI | 39.6 | 52.1 | 53.4 |
231
+ | [small](https://huggingface.co/KBLab/kb-whisper-small) | **KBLab** | **7.3** | **6.4** | **6.6** |
232
  | | OpenAI | 20.6 | 26.4 | 26.4 |
233
+ | [medium](https://huggingface.co/KBLab/kb-whisper-medium) | **KBLab** | **6.6** | **5.4** | **5.8** |
234
  | | OpenAI | 12.1 | 15.8 | 17.1 |
235
+ | [large-v3](https://huggingface.co/KBLab/kb-whisper-large) | **KBLab** | **5.4** | **4.1** | **5.2** |
236
  | | OpenAI | 7.8 | 9.5 | 11.3 |
237
 
238
+ #### WER for different KBLab stage2 versions
239
 
240
+ | Model size | | FLEURS | CommonVoice | NST |
241
+ |------------|---------|--------|-------------|------|
242
+ | [tiny](https://huggingface.co/KBLab/kb-whisper-tiny) | **standard** | **13.2** | **12.9** | **11.2** |
243
+ | | strict | 14.1 | 13.4 | 11.0 |
244
+ | | subtitle | 13.3 | 12.9 | 11.4 |
245
+ | [base](https://huggingface.co/KBLab/kb-whisper-base) | **standard** | **9.1** | **8.7** | **7.8** |
246
+ | | strict | 10.4 | 9.6 | 8.4 |
247
+ | | subtitle | 9.1 | 8.7 | 7.9 |
248
+ | [small](https://huggingface.co/KBLab/kb-whisper-small) | **standard** | **7.3** | **6.4** | **6.6** |
249
+ | | strict | 8.2 | 7.0 | 6.7 |
250
+ | | subtitle | 7.3 | 6.4 | 6.6 |
251
+ | [medium](https://huggingface.co/KBLab/kb-whisper-medium) | **standard** | **6.6** | **5.4** | **5.8** |
252
+ | | strict | 6.8 | 5.4 | 6.0 |
253
+ | [large-v3](https://huggingface.co/KBLab/kb-whisper-large) | **standard** | **5.4** | **4.1** | **5.2** |
254
+ | | strict | 5.3 | 4.0 | 5.1 |
255
+
256
+ #### BLEU Score compared to OpenAI
257
  | Model size | | FLEURS | CommonVoice | NST |
258
  |------------|---------|--------|-------------|------|
259
+ | [tiny](https://huggingface.co/KBLab/kb-whisper-tiny) | **KBLab** | **76.6** | **73.7** | **74.3** |
260
  | | OpenAI | 26.9 | 21.1 | 24.0 |
261
+ | [base](https://huggingface.co/KBLab/kb-whisper-base) | **KBLab** | **83.2** | **79.9** | **78.3** |
262
  | | OpenAI | 41.1 | 32.5 | 36.9 |
263
+ | [small](https://huggingface.co/KBLab/kb-whisper-small) | **KBLab** | **86.6** | **83.5** | **79.6** |
264
  | | OpenAI | 64.0 | 56.5 | 58.2 |
265
+ | [medium](https://huggingface.co/KBLab/kb-whisper-medium) | **KBLab** | **87.6** | **85.0** | **80.2** |
266
  | | OpenAI | 77.1 | 70.1 | 68.9 |
267
+ | [large-v3](https://huggingface.co/KBLab/kb-whisper-large) | **KBLab** | **89.8** | **87.2** | **81.1** |
268
  | | OpenAI | 84.9 | 79.1 | 75.1 |
269
 
270
+ #### BLEU Score for different KBLab stage2 versions
271
+ | Model size | | FLEURS | CommonVoice | NST |
272
+ |------------|---------|--------|-------------|------|
273
+ | [tiny](https://huggingface.co/KBLab/kb-whisper-tiny) | **standard** | **76.6** | **73.7** | **74.3** |
274
+ | | strict | 75.3 | 72.9 | 74.6 |
275
+ | | subtitle | 76.6 | 73.7 | 74.1 |
276
+ | [base](https://huggingface.co/KBLab/kb-whisper-base) | **standard** | **83.2** | **79.9** | **78.3** |
277
+ | | strict | 81.0 | 78.4 | 77.5 |
278
+ | | subtitle | 83.2 | 79.8 | 78.2 |
279
+ | [small](https://huggingface.co/KBLab/kb-whisper-small) | **standard** | **86.6** | **83.5** | **79.6** |
280
+ | | strict | 84.9 | 82.4 | 79.3 |
281
+ | | subtitle | 86.6 | 83.5 | 79.6 |
282
+ | [medium](https://huggingface.co/KBLab/kb-whisper-medium) | **standard** | **87.6** | **85.0** | **80.2** |
283
+ | | strict | 87.3 | 84.9 | 80.1 |
284
+ | [large-v3](https://huggingface.co/KBLab/kb-whisper-large) | **standard** | **89.8** | **87.2** | **81.1** |
285
+ | | strict | 90.0 | 87.4 | 81.2 |
286
+
287
+
288
 
289
  ### Acknowledgements
290