# Chuvash ASR (Aisar) — SpeechCollector

Three open Whisper models fine-tuned for Chuvash (ISO 639-1: cv), plus the
largest open collection of Chuvash speech data.
Site: https://chuvash-asr.vercel.app (ru / en / cv).

## Models (test = Common Voice Chuvash, 1,288 utterances, beam-5)

| Model | Base | Params | WER ortho | CER ortho | WER norm | Notes |
|---|---|---|---|---|---|---|
| Aisar-small  | openai/whisper-small           | 244M | 20.3%  | 5.35% | 15.4%  | lightest, ~0.49 GB; est. ~6.5x realtime* |
| Aisar-medium | openai/whisper-medium          | 769M | 17.28% | 3.83% | 12.36% | balanced, ~1.5 GB; est. ~3.3x realtime* |
| Aisar-turbo  | bond005/whisper-podlodka-turbo | 809M | 16.02% | 3.63% | 11.13% | best AND fastest; measured ~13x realtime beam-5 on V100 fp32 (~22x greedy) |

*Estimates from OpenAI's official relative speeds (small ~4x, medium ~2x,
turbo ~8x vs large) anchored to the measured turbo RTF. Turbo is faster than
small/medium despite its size: it has only 4 decoder layers.

"ortho" = orthographic (casing + punctuation counted). "norm" = after
Whisper-style text normalization (lowercase, ё→е, punctuation stripped).

## Which model to pick

- Default: Aisar-turbo — both the most accurate and the fastest.
- Low memory / smallest checkpoint (~0.49 GB): Aisar-small.
- Middle ground on memory: Aisar-medium.

## How to run (transformers)

pipeline("automatic-speech-recognition", model=<repo>, chunk_length_s=30,
stride_length_s=(5, 5)) with generate_kwargs={"language": "russian",
"task": "transcribe", "num_beams": 5}. Input any sample rate (resampled to
16 kHz). Chuvash has no token in the Whisper vocabulary — "russian" works
best. num_beams=1 is ~1.7x faster at ~+0.3pp normalized WER (turbo). For
long spontaneous audio add temperature fallback: temperature=(0.0, 0.2, 0.4,
0.6, 0.8, 1.0), compression_ratio_threshold=2.4, logprob_threshold=-1.0,
no_speech_threshold=0.6.

Repos (private until release):
- https://huggingface.co/SpeechCollector/whisper-chuvash-small
- https://huggingface.co/SpeechCollector/whisper-chuvash-medium
- https://huggingface.co/SpeechCollector/whisper-chuvash-turbo

## Datasets (https://huggingface.co/SpeechCollector)

- ChuvashAsrDataset — 36,781 utt / 49.5 h labeled. Splits: train 34,251 /
  val 1,242 / test 1,288 (val+test are Common Voice only). This is the
  training corpus of all three models. Sources: alexantonov/chuvash_voice
  (38.7 h, CC0), Common Voice 25.0 (5.6 h, CC0), ftyers Turkic_TTS
  (5.2 h, CC-BY-SA).
- ChuvashRaw — 370,976 seg / 219.0 h (audiobooks, Bible, lyrics), CC BY-NC 4.0
- ChuvashConversationRaw — 61,020 seg / 125.3 h (conversations), CC BY-NC 4.0
- ChuvashNewsRaw — 255,326 seg / 735.8 h (news/broadcast), CC BY-NC 4.0

Raw corpora total ≈1,080 h / 687,322 segments. Their `text` fields are
pseudo-labels generated by Aisar-turbo; this data was NOT used to train the
models (evaluation is leak-free). Long recordings were segmented with Silero
VAD (0.5–60 s chunks, FLAC). Caveat: recordings contain some Russian speech —
filter for Chuvash-only tasks. An MMS-LID + binary ru/cv classifier check on
ChuvashRaw: of ~297 h analyzed, ~256 h (~86%) passed score_chv >= 0.5.

## Links

- Org: https://huggingface.co/SpeechCollector
- Site: https://chuvash-asr.vercel.app
- Contact: https://t.me/Michaelya
- Paper: coming soon on arXiv