Distil-Whisper Large v3.5

Distil-Whisper is Hugging Face's knowledge-distilled version of Whisper, introduced in Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling (Gandhi, von Platen & Rush, 2023). The v3.5 release is the latest English checkpoint in the family.

• Encoder–decoder Transformer, encoder copied & frozen from Whisper Large v3
• Only 2 decoder layers (vs 32 in Whisper Large v3)
• Trained on 98k hours of diverse public speech (4× more than prior Distil-Whisper), pseudo-labelled by Whisper Large v3
• 'Patient' teacher with extended schedule + aggressive SpecAugment

• Drop-in replacement for Whisper Large v3 on English audio
• ~1.5× faster than Whisper-Large-v3-Turbo; ~2× faster than Whisper-Large-v3
• Supports both chunked and OpenAI sequential long-form transcription algorithms
• Works as a perfect speculative decoding draft model for Whisper Large v3
• Compatible with Whisper.cpp, Faster-Whisper, Transformers, Transformers.js, MLX

Production English transcription, on-device / browser ASR via ONNX, speculative decoding to accelerate Whisper, long-form podcast/meeting transcription.