README.md

---
language:
  - cv
license: cc-by-nc-4.0
library_name: transformers
pipeline_tag: automatic-speech-recognition
base_model: bond005/whisper-podlodka-turbo
datasets:
  - SpeechCollector/ChuvashAsrDataset
tags:
  - whisper
  - chuvash
  - chv
  - speech-recognition
  - asr
metrics:
  - wer
  - cer
model-index:
  - name: whisper-chuvash-turbo
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        dataset:
          type: SpeechCollector/ChuvashAsrDataset
          name: ChuvashAsrDataset (Common Voice test split)
          split: test
        metrics:
          - type: wer
            value: 16.02
            name: WER (orthographic, beam-5)
          - type: cer
            value: 3.63
            name: CER (orthographic, beam-5)
          - type: wer
            value: 11.13
            name: WER (Whisper-normalized, beam-5)
---

# whisper-chuvash-turbo

Whisper fine-tuned for **Chuvash** (`chv`) — the most accurate — and the fastest — model of the family in the **Aisar**
family by [SpeechCollector](https://huggingface.co/SpeechCollector). The
family was trained on [ChuvashAsrDataset](https://huggingface.co/datasets/SpeechCollector/ChuvashAsrDataset)
(~49.5 h of labeled Chuvash speech) and evaluated on its Common Voice test
split (1,288 utterances).

- **Base model:** [bond005/whisper-podlodka-turbo](https://huggingface.co/bond005/whisper-podlodka-turbo) (809M parameters)
- **WER / CER (orthographic, beam-5):** 16.02% / 3.63%
- **WER / CER (Whisper-normalized):** 11.13% / 2.6%
- **Checkpoint size:** ~1.6 GB
- **Inference speed:** ~13x real time (measured, beam-5, V100 fp32; ~22x greedy)

The model was initialized from [bond005/whisper-podlodka-turbo](https://huggingface.co/bond005/whisper-podlodka-turbo),
a Russian-adapted large-v3-turbo. That starting point pays off for Chuvash,
which carries many Russian loanwords. Despite being the largest model of the
family by parameter count, it is also the fastest: large-v3-turbo keeps only
4 decoder layers.

Beyond recognition, this model produced the pseudo-labels for the raw corpora
([ChuvashRaw](https://huggingface.co/datasets/SpeechCollector/ChuvashRaw),
[ChuvashConversationRaw](https://huggingface.co/datasets/SpeechCollector/ChuvashConversationRaw),
[ChuvashNewsRaw](https://huggingface.co/datasets/SpeechCollector/ChuvashNewsRaw)
— about 1,080 hours in total). None of that data was used to train the models,
so the evaluation is leak-free.

## The Aisar family

| Model | Base | Params | WER | CER | WER (norm.) |
|---|---|---|---|---|---|
| [whisper-chuvash-small](https://huggingface.co/SpeechCollector/whisper-chuvash-small) | openai/whisper-small | 244M | 20.3% | 5.35% | 15.4% |
| [whisper-chuvash-medium](https://huggingface.co/SpeechCollector/whisper-chuvash-medium) | openai/whisper-medium | 769M | 17.28% | 3.83% | 12.36% |
| [whisper-chuvash-turbo](https://huggingface.co/SpeechCollector/whisper-chuvash-turbo) | bond005/whisper-podlodka-turbo | 809M | **16.02%** | **3.63%** | **11.13%** |

Orthographic = text compared as-is, with casing and punctuation.
Normalized = same beam-5 run scored after Whisper-style normalization
(lowercase, ё→е, punctuation stripped). Speed estimates for small/medium are
derived from OpenAI's official relative speeds (small ~4x, medium ~2x,
turbo ~8x vs large) anchored to the measured turbo RTF.

## How to use

Chuvash has no language token in the Whisper vocabulary, so generation uses
`language="russian"` — this is intentional and gives the best results. The
numbers above were obtained with beam-5; `num_beams=1` is ~1.7x faster at a
small quality cost (~+0.3 pp normalized WER for turbo).

```python
import torch
from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="SpeechCollector/whisper-chuvash-turbo",
    torch_dtype=torch.float16,   # use torch.float32 on CPU
    device="cuda:0",             # or "cpu"
    chunk_length_s=30,           # handles audio longer than 30 s
    stride_length_s=(5, 5),
)

out = asr(
    "speech.wav",  # any sample rate — the pipeline resamples to 16 kHz
    generate_kwargs={
        "language": "russian",
        "task": "transcribe",
        "num_beams": 5,
    },
)
print(out["text"])
```

For long spontaneous audio (conversations, broadcast), add a temperature
fallback to `generate_kwargs` — it tames repetition loops on hard segments:

```python
"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,
```

## Training data

[ChuvashAsrDataset](https://huggingface.co/datasets/SpeechCollector/ChuvashAsrDataset):
36,781 utterances / ~49.5 h, 16 kHz mono. 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). Splits: train 34,251 / val 1,242 / test 1,288 — validation and
test contain Common Voice recordings only.

## Limitations

- Evaluated on read speech (Common Voice); expect higher error rates on
  spontaneous or noisy audio.
- Fine-tuned almost exclusively on monolingual Chuvash, so the model performs
  best on Chuvash-dominant audio and degrades on Chuvash–Russian
  code-switched speech. Bilingual fine-tuning is a natural next step.
- Output is orthographic Chuvash (ӑ, ӗ, ҫ, ӳ); casing and punctuation come
  from the model and are not guaranteed.

## License and acceptable use

- **License:** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — free for research and other non-commercial use.
- **Commercial use:** possible under a separate agreement with the SpeechCollector community — reach out on Telegram: [@Michaelya](https://t.me/Michaelya).
- **Surveillance:** please do not use this model to build systems for surveillance, covert monitoring, or tracking of individuals.

## Links

- Project website (ru / en / cv): https://chuvash-asr.vercel.app
- Datasets and models: https://huggingface.co/SpeechCollector
- Contact: [@Michaelya](https://t.me/Michaelya)
- Paper: coming soon on arXiv