Distil-Whisper

Distil-Whisper is a family of speech recognition models for developers who want to transcribe English audio on their own hardware. The models build on Whisper, reducing model size and processing time with some differences in accuracy.

• English transcription: Transcribe short clips and long recordings with distilled models that perform close to Whisper in reported evaluations.
• Long recordings: distil-large-v3 supports sequential processing for more accurate transcription and chunked processing for faster results on a single large file.
• Noisy audio: Reported tests show good performance with background noise, fewer repeated word sequences, and fewer inserted errors than Whisper.
• Whisper acceleration: Use Distil-Whisper as an assistant model through speculative decoding, which is designed to preserve Whisper's output while reducing inference time.

The models work with Hugging Face Transformers, and distil-large-v3 is compatible across Whisper libraries. GPU acceleration can use Flash Attention on compatible hardware. The open-source project uses the MIT license, which permits commercial use.

The supplied models recognize English only. Training code supports distilling Whisper for other languages, but that requires training a separate model.