distil-whisper

Why we included this project

Distil-Whisper is a practical answer to the biggest complaint about OpenAI's Whisper: it is too slow and heavy for high-volume or real-time transcription. The distilled checkpoints stay within about 1% word error rate of Whisper large-v3 on out-of-distribution sets while running roughly six times faster and cutting the model size in half, so the savings are real rather than theoretical. You load them through the standard Hugging Face Transformers automatic speech recognition pipeline, exactly as you would with the original model, and the README walks through short-form clips, local audio files, timestamps, and beam search. If memory is tight, the distil-small.en variant drops to 166M parameters at a modest accuracy cost, which makes it a reasonable fit for on-device or mobile work. One boundary to plan around: these checkpoints are English-only, so multilingual teams should look at Whisper Turbo instead.