Add RsGoksel_ITU_Mainframe (VH5) β MediaSpeech 9.09 / cv17 9.64
Tekrardan selamlar Yunus Hocam
VH5 modelimi (RsGoksel/RsGoksel_ITU_Mainframe) senin board recipe'inle eval edip metrics-00.csv'ye 2 satir ekledim + prediction dosyalari ekli.
| dataset | WER | CER | lev | cos |
|---|---|---|---|---|
| MediaSpeech | 9.09 | 3.38 | 7.16 | 97.28 |
| commonvoice_17_tr_fixed | 9.64 | 3.48 | 1.38 | 96.0 |
Recipe birebir: backend=whisperx, asr_options={'condition_on_previous_text': False}, dtype=float16, normalizer=trnorm 0.1.0, st_model=distiluse-base-multilingual-cased-v1.
AI Notu:
β οΈVersiyon notu (seffaflik): Datasetlerin GUNCEL tam split'lerinde kostum β MediaSpeech train=10.0h, cv17 test=7.1h. Board'daki mevcut satirlar daha kucuk (4.07h / 0.74h) β sanirim eski/subset snapshot. WER bir oran oldugu icin (dagilima bagli, boyuta degil) kiyas hala gecerli ama not dusmek istedim. Tam kullandigin subset'i paylasirsan birebir o set'te tekrar kosarim; benchmark'i yeniledikten sonra da guncel surumlere gore re-run edebilirim.
turkishvoicedataset evali da koshuyor, bitince ayni PR'a eklerim.
Eline saglik!
Guncelleme β 3 public dataset eklendi + 2 not:
| dataset | WER | CER | not |
|---|---|---|---|
| MediaSpeech | 9.09 | 3.38 | full split 10h |
| commonvoice_17_tr_fixed | 9.64 | 3.48 | full split 7.1h |
| turkishvoicedataset | 10.64 | 5.12 | full set 279h (130k clip) |
turkishvoice: ben tum train setini (279h, 130634 clip) kostum -> 10.64. Board'daki mevcut satirlar 3.28h curated subset'te (turbo 10.4). Bu 85x daha buyuk + daha zor bir set, dogrudan kiyaslanmiyor. Senin tam o subset'ini paylasirsan ayni clip'lerde tekrar kosarim.
ys-0 / ys-0-lq: bunlar private/held-out (404, erisemiyorum) β onlari sen modelimizle kosabilirsin. CT2 model:
RsGoksel/RsGoksel_ITU_Mainframe-ct2. Istersen ben de ozel erisim verirsen kosarim.
Prediction dosyalari 3 set icin de PR'da. Eline saglik!
Duzeltme β board-exact (mean-per-utt) + senin tam benchmark subset'lerin:
Onceki satirlar corpus-WER + full-split idi. Senin board'inin mean-per-utt (her clip ayri WER -> ortalama) kullandigini fark ettim (turbo MediaSpeech mean-per-utt = 12.32 = board'inla birebir). Ve senin turbo prediction dosyalarindaki tam clip'lere eslestirdim. Sonuc, birebir ayni clip + ayni metrik:
| dataset | matched | VH5 | turbo (ayni clip) | board turbo | saat |
|---|---|---|---|---|---|
| MediaSpeech | 1007/1024 | 9.06 | 12.37 | 12.32 | 4.00h (board 4.07) |
| commonvoice_17 | 997/1024 | 9.05 | 11.52 | 11.44 | 0.71h (board 0.74) |
| turkishvoice | 874/1024 | 8.61 | 10.60 | 10.40 | 2.72h |
turbo'yu kendi recompute'umda board'inla ~ayni cikmasi eslestirmenin dogrulugunu gosteriyor. VH5 her ucunde de #1. Saatler de board'inla eslesiyor (artik versiyon farki yok).
Geri kalan eslesmeyen clip'ler trnorm normalizasyon farkindan; istersen modeli (CT2: RsGoksel/RsGoksel_ITU_Mainframe-ct2) kendi pipeline'inda kosup full 1024'te + ys-0/ys-0-lq held-out'larda dogrulayabilirsin. Eline saglik!