Three models

One corpus, three Whisper sizes. All numbers come from the Common Voice test set: 1,288 utterances, beam-5, orthographic WER and CER. Click a card to open the model page.

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

What's released

3 modelsWhisper small / medium / turbo, fine-tuned for Chuvash
16.02%best WER (turbo); 11.13% after normalization
49.5 hlabeled speech from open sources — the training corpus
1,080 hadditionally auto-labeled by the turbo model

Everything is open on HuggingFace, under the SpeechCollector organization: the models can be downloaded and run. Click a model card for its metrics, a usage snippet, and a link to the repository.