OpenAI Unveils Whisper Turbo: 8x Faster, 2x Smaller, Transcribes 2 Minutes in 12 Seconds

Oct 1, 2024, 08:54 AM

On October 1, 2024, OpenAI released Whisper Turbo, a new version of its Whisper speech-to-text model. The optimized model offers 8 times faster transcription speed and uses 2 times fewer parameters than Whisper Large v3, with minimal degradation in accuracy. Whisper Turbo features a reduced model size of 809 million parameters and utilizes 4 decoder layers instead of 32. It is multilingual, supports both word and chunk-level timestamps, and is 40% more VRAM efficient. The model can transcribe two minutes of audio in approximately 12 seconds and can run locally in a browser using Transformers.js, making it accessible for applications with limited resources.