insanely-fast-whisper
A local audio transcription CLI runs Whisper and Distil-Whisper on CUDA GPUs or MPS-enabled Macs, with optional Flash Attention 2 and Apache 2.0 licensing.
insanely-fast-whisper is a command-line audio transcription tool for people who want to process recordings on their own hardware. It focuses on reducing transcription time through GPU acceleration.
- Whisper model choice: Transcribe audio with OpenAI's Whisper large-v3 or Distil-Whisper, including the distil-large-v2 model used in the project's benchmarks.
- Accelerated transcription: GPU batching and optional Flash Attention 2 speed up processing through Hugging Face Transformers and Optimum.
The CLI supports CUDA devices and MPS-enabled Macs, including CUDA use on Windows. Transcription runs on-device. The Mac MPS backend uses more memory than CUDA, which can affect how much audio you can process at once.
The published speed benchmarks use an NVIDIA A100, so those timings don't establish expected performance on a personal computer. The project is open source under Apache 2.0.
Logotyper
