Whisper Large V2 French
| Property | Value |
|---|---|
| Parameter Count | 1.55B parameters |
| License | Apache 2.0 |
| Language | French |
| Architecture | Sequence-to-sequence |
What is whisper-large-v2-french?
This is a specialized French Automatic Speech Recognition (ASR) model fine-tuned from OpenAI's Whisper Large V2. It has been optimized using over 2,200 hours of French speech data from diverse sources including Common Voice, LibriSpeech, VoxPopuli, and other datasets. The model demonstrates exceptional performance with Word Error Rates (WER) as low as 4.03% on MLS and 4.98% on Fleurs datasets.
Implementation Details
The model builds upon the Whisper architecture, requiring 16kHz audio input and notably does not predict casing or punctuation. It leverages the Transformers library and provides both greedy and beam search decoding options for optimal transcription results.
- Trained on 7 major French speech datasets
- Supports both greedy and beam search (width=5) inference
- Requires 16kHz audio sampling rate
- Implements forced decoder IDs for French language
Core Capabilities
- State-of-the-art WER performance across multiple benchmarks
- Handles various French accents including African French
- Efficient transcription with both fast (greedy) and accurate (beam search) options
- Seamless integration with Hugging Face's pipeline API
Frequently Asked Questions
Q: What makes this model unique?
This model represents a significant improvement over the base Whisper model for French ASR, with WER reductions of up to 50% compared to the original model. It's been extensively trained on diverse French speech data, making it particularly robust for real-world applications.
Q: What are the recommended use cases?
The model is ideal for French speech transcription tasks requiring high accuracy, such as media transcription, subtitling, and voice command systems. It's particularly effective for applications needing to handle various French accents and speech patterns.
