faster-whisper-tiny-eu-cv22.0-int8

Basque (eu) speech recognition model in CTranslate2 format (int8 quantized), ready to use with faster-whisper and the Home Assistant Whisper app.

Origin

Base model	openai/whisper-tiny
Basque fine-tune	zuazo/whisper-tiny-eu-cv22.0 by Xabier de Zuazo
Training data	Common Voice 22.0 (eu)
Reported WER	see source model card
Size	41 MB (int8)
License	MIT (inherited from the source model)

This is a format and quantization conversion only: the weights are unchanged, converted to CTranslate2 and quantized to int8. It is distributed under the same MIT license as the source model.

How it was created

ct2-transformers-converter \
  --model zuazo/whisper-tiny-eu-cv22.0 \
  --output_dir faster-whisper-tiny-eu-cv22.0-int8 \
  --quantization int8 \
  --copy_files tokenizer.json preprocessor_config.json generation_config.json

Usage

faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("faster-whisper-tiny-eu-cv22.0-int8", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.wav", language="eu", beam_size=5)
print(info.language, info.language_probability)
print(" ".join(segment.text for segment in segments))

Home Assistant — Whisper app

Settings → Apps → Whisper → Configuration:

language: eu
model: custom
custom_model: itzune/faster-whisper-tiny-eu-cv22.0-int8

Then set the Assist pipeline language to eu. This repository must stay public: the Whisper app does not authenticate to the Hub.

Notes

• Pin language: eu. Whisper's auto detection is much slower.
• Runs on CPU; int8 keeps memory usage low. CPUs with AVX2 are considerably faster.
• Lightest option; slightly lower accuracy on compound words.

Credits

Xabier de Zuazo (zuazo) for the Basque fine-tunes · OpenAI for Whisper · SYSTRAN/OpenNMT for CTranslate2 and faster-whisper.