kyrgyz-whisper-small

Kyrgyz speech recognition (кыргызча кепти текстке айландыруу) with Whisper-small.

🛠️ Toolkit: github.com/ramput-labs/kyrgyz-whisper-small

CLI, Python API, long-audio transcription, microphone input, web demo and WER/CER evaluation.

git clone https://github.com/ramput-labs/kyrgyz-whisper-small.git
cd kyrgyz-whisper-small && make quickstart

Usage with transformers

import soundfile as sf
from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="ramput-labs/kyrgyz-whisper-small")
audio, sr = sf.read("audio.wav", dtype="float32")
print(asr({"raw": audio, "sampling_rate": sr}, generate_kwargs={"language": "kk", "task": "transcribe"})["text"])

Whisper has no Kyrgyz token, so pass language="kk"; it gives the best accuracy. This snippet handles clips up to 30 s. For longer audio, use the toolkit.

Results

FLEURS Kyrgyz dev, 100 clips without digits: WER 16.3% / CER 4.5%. Numbers are written as words ("жети миң"), so references with digits score lower.

Credits

• Weights: the-cramer-project/AkylAI-STT-small by The Cramer Project, a Kyrgyz fine-tune of openai/whisper-small by OpenAI. We did not retrain them; we re-saved them with a Kyrgyz generation setup (language="kk") and added evaluation results.
• Evaluation audio: google/fleurs (CC-BY-4.0).

License

Apache-2.0 (see LICENSE), inherited from AkylAI-STT-small and Whisper. The toolkit code is MIT.