加载中…
加载中…
Fine-tuned Moonshine ASR model for French language
This is a fine-tuned version of UsefulSensors/moonshine-tiny specifically optimized for French speech recognition. The model achieves state-of-the-art performance for its size (27M parameters) on French ASR tasks.
Links:
pip install --upgrade pip
pip install --upgrade transformers datasets[audio]
from transformers import MoonshineForConditionalGeneration, AutoProcessor
import torch
import torchaudio
# Load model and processor
model = MoonshineForConditionalGeneration.from_pretrained('Cornebidouil/moonshine-tiny-fr')
processor = AutoProcessor.from_pretrained('Cornebidouil/moonshine-tiny-fr')
# Load and resample audio to 16kHz
audio, sr = torchaudio.load("french_audio.wav")
if sr != 16000:
audio = torchaudio.functional.resample(audio, sr, 16000)
audio = audio[0].numpy() # Convert to mono
# Prepare inputs
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
# Generate transcription
# Calculate max_new_tokens to avoid truncation (5 tokens per second is optimal for French)
audio_duration = len(audio) / 16000
max_new_tokens = int(audio_duration * 5)
generated_ids = model.generate(**inputs, max_new_tokens=max_new_tokens)
transcription = processor.decode(generated_ids[0], skip_special_tokens=True)
print(transcription)
For production deployments with:
See the included inference.py script in the fine-tuning guide.
Moonshine is a compact sequence-to-sequence ASR model designed for efficient on-device inference:
Evaluated on Multilingual LibriSpeech (MLS) French test set:
| Metric | Score |
|---|---|
| Word Error Rate (WER) | 21.8% |
| Character Error Rate (CER) | ~10% |
| Real-Time Factor (RTF) | 0.11x (CPU) |
Inference Speed: ~9x faster than real-time on CPU, enabling live transcription.
| Model | Size | Language | WER (MLS-FR) |
|---|---|---|---|
| Whisper-tiny | 39M | Multilingual | ~25% |
| Moonshine-tiny-fr | 27M | French | 21.8% |
| Whisper-base | 74M | Multilingual | ~18% |
Moonshine-tiny-fr achieves competitive performance with 30% fewer parameters than Whisper-tiny. While being a proof of concept. More work should be done to create a proper and robust dataset.
Please refer to my Github repo for the training procedure :
✅ French Speech Recognition
✅ Resource-Constrained Environments
✅ Hogwarts Legacy SpellCaster
Pierre Chéneau (Cornebidouil)
Geologist, Developer and maintainer of this fine-tuned French model.
Links:
@misc{cheneau2026moonshine-tiny-fr,
author = {Pierre Chéneau (Cornebidouil)},
title = {Moonshine-Tiny-FR: Fine-tuned French Speech Recognition},
year = {2026},
publisher = {HuggingFace},
url = {https://huggingface.co/Cornebidouil/moonshine-tiny-fr}
}
@misc{cheneau2026moonshine-finetune,
author = {Pierre Chéneau (Cornebidouil)},
title = {Moonshine ASR Fine-Tuning Guide},
year = {2026},
publisher = {GitHub},
url = {https://github.com/pierre-cheneau/finetune-moonshine-asr}
}
@misc{jeffries2024moonshinespeechrecognitionlive,
title={Moonshine: Speech Recognition for Live Transcription and Voice Commands},
author={Nat Jeffries and Evan King and Manjunath Kudlur and Guy Nicholson and James Wang and Pete Warden},
year={2024},
eprint={2410.15608},
archivePrefix={arXiv},
primaryClass={cs.SD},
url={https://arxiv.org/abs/2410.15608},
}
@inproceedings{panayotov2015librispeech,
title={Multilingual LibriSpeech: A Corpus for Speech Recognition in Multiple Languages},
author={Pratap, Vineel and Xu, Qiantong and Sriram, Anuroop and Synnaeve, Gabriel and Collobert, Ronan},
booktitle={Interspeech},
year={2020}
}
This model is released under the MIT License, consistent with the base Moonshine model.
MIT License
Copyright (c) 2026 Pierre Chéneau (Cornebidouil)
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction...
Questions? Open an issue on the fine-tuning guide repository or check the documentation.
Want to fine-tune for your language? See the complete fine-tuning guide.