distil-large-v3.5

Distil-Whisper is the knowledge-distilled version of OpenAI's Whisper-Large-v3, described in the paper Robust Knowledge Distillation via Large-Scale Pseudo Labelling. As the newest addition to the Distil-Whisper English family, Distil-Large-v3.5 maintains the high efficiency of its predecessors while delivering better performance. Compared to earlier models, Distil-Large-v3.5 has been trained on over 4× more diverse public data (98k hours) and uses a "patient" teacher with an extended training schedule and aggressive data augmentation (SpecAugment) during distillation. This results in enhanced robustness and accuracy compared to previous Distil-Whisper models, making it suitable as a drop-in replacement. 1. It offers a different balance between accuracy and efficiency, remains **~1.5x faster** than Whisper-Large-v3-Turbo while performing slightly better on short-form transcription and falling ~1% behind on long-form transcription. 2. It works perfectly as a draft model for **speculative decoding** with Whisper-Large-v3. By keeping the encoder frozen during training, we need to load just two extra decoder layers and forward the encoder only once. This achieves ~2x faster inference compared to Whisper-Large-v3 while maintaining identical outputs.

번역 준비 중입니다. 원문을 표시합니다.