whisper-small-fr-cv-100k

Whisper Small French - Fine-tuned on Common Voice

This model is a fine-tuned version of openai/whisper-small on French speech data from Common Voice.

Model Description

• Base Model: OpenAI Whisper Small (244M parameters)
• Language: French
• Task: Automatic Speech Recognition (Transcription)
• Training Data: Common Voice 13.0 French dataset
• Training Samples: 100,000 samples
• Training Duration: 3 epochs

Training Details

Training Hyperparameters

• Optimizer: AdamW
• Learning Rate: 1e-5
• Batch Size: 8
• Gradient Accumulation: 2 steps
• Mixed Precision: FP16
• Scheduler: Cosine Annealing
• Max Epochs: 3
• Early Stopping: Patience of 2 epochs on validation WER

Hardware

• GPU: NVIDIA T4/L4
• Training Time: ~4 hours

Usage

from transformers import pipeline

# Load the model
pipe = pipeline(
    "automatic-speech-recognition",
    model="keypa/whisper-small-fr-cv-100k",
    device=0  # Use GPU (or -1 for CPU)
)

# Transcribe audio
result = pipe("path/to/your/french/audio.wav")
print(result["text"])

Or with the model and processor directly:

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

# Load model and processor
processor = WhisperProcessor.from_pretrained("keypa/whisper-small-fr-cv-100k")
model = WhisperForConditionalGeneration.from_pretrained("keypa/whisper-small-fr-cv-100k")
model.to("cuda")  # or "cpu"

# Load audio
import librosa
audio, sr = librosa.load("audio.wav", sr=16000)

# Process
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
input_features = input_features.to("cuda")

# Generate
predicted_ids = model.generate(input_features)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription[0])

Performance

This model was trained on 100k samples and achieves good performance on French speech recognition tasks.

Limitations

• Optimized for French language only
• Performance may vary on accents not well-represented in Common Voice
• Best suited for clear audio recordings

Citation

If you use this model, please cite:

@misc{whisper-small-fr-100k,
  author = {keypa},
  title = {Whisper Small French Fine-tuned},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/keypa/whisper-small-fr-cv-100k}}
}

Acknowledgements

• Base model: OpenAI Whisper
• Training data: Mozilla Common Voice