Insanely Fast Whisper

Insanely Fast Whisper wraps optimized Whisper inference in an opinionated command-line tool for on-device transcription. It combines Hugging Face Transformers, Optimum, and Flash Attention 2 with fp16 precision and batched inference, batch size 24 by default, to transcribe about 150 minutes of audio in roughly 98 seconds with Whisper Large v3 on an NVIDIA A100. The CLI supports OpenAI Whisper checkpoints and the distilled distil-whisper variants, handles both transcription and translation, produces chunk-level or word-level timestamps, and adds speaker diarization through Pyannote.audio. It runs on CUDA GPUs and on Apple Silicon via MPS, installs with pipx, and accepts a file path or URL as input. The project began as a benchmarking showcase for how fast Whisper could run with standard Transformers optimizations and grew into a community-maintained tool. Released under the Apache 2.0 license, it suits developers who want fast local transcription without writing pipeline code.