internetoftim commited on
Commit
79cfd2e
·
verified ·
1 Parent(s): 44b9448

Update dataset card (pld.md)

Browse files
Files changed (1) hide show
  1. README.md +35 -3
README.md CHANGED
@@ -219,16 +219,48 @@ same-sentence cross-speaker pairs mined from PLD's shared prompt lists (the
219
  corpus has no parallel translations, but many speakers reading the same prompt
220
  is exactly the parallel data voice conversion needs).
221
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
222
  Reproduce any of them in one command:
223
 
224
  ```bash
225
  python finetune_tts.py --dataset pld --language ceb --push
 
226
  python finetune_s2s.py --push
227
  ```
228
 
229
- An ASR baseline (whisper-small per language) has **not** been trained yet — a
230
- Bikol or Cebuano one would be the first of its kind in public. If you train
231
- something on PLD, tag this dataset in your model card and it will appear here.
232
 
233
  ---
234
 
 
219
  corpus has no parallel translations, but many speakers reading the same prompt
220
  is exactly the parallel data voice conversion needs).
221
 
222
+ ### Speech recognition — one model per language
223
+
224
+ `openai/whisper-small` finetuned per language, scored on a held-out split.
225
+ **CER is the selection metric**, not WER: Philippine-language orthography
226
+ varies at the word level (hyphenation and affix spelling especially), which
227
+ punishes WER for differences a reader would not consider errors.
228
+
229
+ | Language | Model | WER | CER |
230
+ |---|---|---|---|
231
+ | English (PH) | [`whisper-small-pld-eng`](https://huggingface.co/sapinsapin/whisper-small-pld-eng) | 5.9% | 3.1% |
232
+ | Cebuano | [`whisper-small-pld-ceb`](https://huggingface.co/sapinsapin/whisper-small-pld-ceb) | 11.1% | 4.5% |
233
+ | Filipino | [`whisper-small-pld-fil`](https://huggingface.co/sapinsapin/whisper-small-pld-fil) | 11.5% | 4.6% |
234
+ | Tausug | [`whisper-small-pld-tsg`](https://huggingface.co/sapinsapin/whisper-small-pld-tsg) | 12.5% | 4.9% |
235
+ | Waray | [`whisper-small-pld-war`](https://huggingface.co/sapinsapin/whisper-small-pld-war) | 15.4% | 7.6% |
236
+ | Hiligaynon | [`whisper-small-pld-hil`](https://huggingface.co/sapinsapin/whisper-small-pld-hil) | 16.1% | 8.0% |
237
+ | Bikol | [`whisper-small-pld-bcl`](https://huggingface.co/sapinsapin/whisper-small-pld-bcl) | 17.9% | 5.6% |
238
+ | Ilocano | [`whisper-small-pld-ilo`](https://huggingface.co/sapinsapin/whisper-small-pld-ilo) | 20.0% | 5.5% |
239
+ | Pangasinan | [`whisper-small-pld-pag`](https://huggingface.co/sapinsapin/whisper-small-pld-pag) | 32.7% | 17.4% |
240
+ | Kapampangan | [`whisper-small-pld-pam`](https://huggingface.co/sapinsapin/whisper-small-pld-pam) | 40.2% | 14.7% |
241
+
242
+ Read these as **in-domain** numbers: PLD is prompted read speech recorded in
243
+ controlled sessions, and the split is random over utterances, so speakers
244
+ overlap between train and test. Expect materially worse performance on
245
+ spontaneous or noisy audio, and re-split on `speaker_id` if you need a
246
+ speaker-disjoint measurement.
247
+
248
+ Two caveats worth stating plainly. Whisper's decoder only has language tokens
249
+ for about 100 languages: English and Tagalog are in the vocabulary, the other
250
+ eight here are not, so they train under the closest token (`<|tl|>`) which
251
+ finetuning repurposes as the language slot. And Pangasinan is data-limited
252
+ (~5.5k utterances total in the corpus), which is the main reason it trails.
253
+
254
  Reproduce any of them in one command:
255
 
256
  ```bash
257
  python finetune_tts.py --dataset pld --language ceb --push
258
+ python finetune_asr.py --dataset pld --language ceb --push
259
  python finetune_s2s.py --push
260
  ```
261
 
262
+ If you train something better on PLD, tag this dataset in your model card and
263
+ it will appear here.
 
264
 
265
  ---
266