Insanely Fast Whisper

Use it when

• Transcribing hours of recorded audio with Whisper large-v3 is far too slow on your GPU and you need batched fp16 or Flash Attention 2 throughput
• You need local transcription with chunk or word timestamps plus optional pyannote speaker diarization from one CLI

What it solves

Not the fit when

• CPU-only machines; the CLI documents support for NVIDIA CUDA GPUs and Apple Silicon MPS
• Live or streaming transcription; it processes files or URLs
• The reported 98-second result used an NVIDIA A100 80GB with specific batching and Flash Attention settings and is not a guarantee for other GPUs
• Diarization requires a Hugging Face token and access to the selected pyannote model
• Only transcribe recordings you are authorized to process
• real-time streaming captions
• cpu-only server transcription
• subtitle file editing workflow

Install

pipx install insanely-fast-whisper

Invoke

insanely-fast-whisper --file-name <file or URL> (add --device-id mps on macOS; --flash True for Flash Attention 2; --hf-token for diarization)

Alternatives

No reviewed alternatives recorded yet.