Research previewчӑваш чӗлхи · Aisar

Three models
that understand
Chuvash speech.

Aisar is three Whisper models fine-tuned on an open labeled corpus of Chuvash speech (49.5 h from three sources, including Common Voice), plus the largest open collection of Chuvash audio — another ≈1,080 hours. Listen to real test-set recordings and compare how small, medium and turbo transcribe live speech.

3 modelssmall · medium · turbo
16.02%best WER — turbo
49.5 hof labeled speech
1,080 hauto-labeled
HuggingFaceDataset▸ Recognition examplesPaper · coming soon on arXiv
▸ sample_017.wav · 4.2s · 16kHz
large-v3-turbo ✓
«Эпир чӑвашла калаҫатпӑр.»
00:00.0000:02.1000:04.20

Three models · one task

Accuracy on the Common Voice test set: 1,288 utterances, beam-5 decoding, orthographic WER — casing and punctuation included.

whisper-small чӑвашла · 244M
light
20.3% WER
0%30%
CER5.35%
WER (norm.)15.4%
Base modelwhisper-small
Checkpoint≈0.49 GB
HuggingFace
whisper-medium чӑвашла · 769M
mid-size
17.28% WER
0%30%
CER3.83%
WER (norm.)12.36%
Base modelwhisper-medium
Checkpoint≈1.5 GB
HuggingFace
whisper-large-v3-turbo чӑвашла · 809M
accurate
16.02% WER
0%30%
CER3.63%
WER (norm.)11.13%
Speed≈13× RT
Checkpoint≈1.6 GB
HuggingFace
1,080hours of audio

The largest open collection of Chuvash speech

Alongside the models we release three datasets: ≈1,080 hours of auto-labeled speech plus a 49.5-hour labeled training corpus (train / val / test). Open Chuvash data at this scale has never been published before.

3 new datasets687,322 segments49.5 h hand-labeled
Explore the data →HuggingFace ↗

Hear it for yourself

real test-set recordings · unedited model outputs
0:00 / 0:00
cv_29078544.mp3all models correct
Reference · Common VoiceЖурналистсем хускатнӑ ыйтусем хушшинче социаллӑ-экономика аталанӑвӗнчен пуҫласа Элтепер сывлӑхӗ таранах пулнӑ.
Aisar-smallЖурналистсем хускатнӑ ыйтусем хушшинче социаллӑ-экономика аталанӑвӗнчен пуҫласа Элтепер сывлӑхӗ таранах пулнӑ.WER 0%
Aisar-mediumЖурналистсем хускатнӑ ыйтусем хушшинче социаллӑ-экономика аталанӑвӗнчен пуҫласа Элтепер сывлӑхӗ таранах пулнӑ.WER 0%
Aisar-turboЖурналистсем хускатнӑ ыйтусем хушшинче социаллӑ-экономика аталанӑвӗнчен пуҫласа Элтепер сывлӑхӗ таранах пулнӑ.WER 0%
0:00 / 0:00
cv_22980135.mp3only turbo got it right
Reference · Common VoiceХӑшне-пӗрне кӑмӑллӑ паллашу, туслӑх, юрату кӗтет.
Aisar-smallКӑшне пӗрне кӑмӑллӑ паллӑ шут туслӑх юрату кӗтет.WER 67%
Aisar-mediumХӑшне пӗрне кӑмӑллӑ паллӑшу туслӑх юрату кӗтет.WER 50%
Aisar-turboХӑшне-пӗрне кӑмӑллӑ паллашу, туслӑх, юрату кӗтет.WER 0%
▸ All recognition examples