Whisper Large v3 Turbo: 8x faster ASR with less than 1% quality degradation

Whisper Large v3 is excellent but slow: on a normal computer without a powerful GPU, transcribing an hour of audio can take an hour or more. For applications requiring fast response — real-time subtitles, voice assistants, instant transcription — it was too slow for much hardware.

OpenAI solved the problem with a technique called "pruning": it analyzed which parts of the model contributed most to quality and which were redundant, then removed about 40% of the decoder parameters (the part that generates text), keeping the encoder (the part that analyzes audio) intact.

The result is a model 8 times faster than Large v3, but making only a tiny number of additional errors — less than 1% increase in error rate for most languages. For practical use, the quality difference is almost imperceptible.

This means high-quality transcription in real time is now possible even on a laptop with a 4GB consumer GPU, or on a server without a dedicated GPU. Professional-quality ASR has become accessible on ordinary hardware.