Insanely Fast Whisper is a command-line tool for transcribing audio files with OpenAI's Whisper model on local hardware.
The tool addresses the slowness of standard Whisper inference by combining several optimization techniques: half-precision floating point arithmetic, batch processing, and Flash Attention 2. These optimizations work together to dramatically reduce transcription time while maintaining accuracy. The tool wraps Whisper models from Hugging Face Transformers and applies these performance enhancements automatically through the CLI.
Developers should choose this tool if they need fast local audio transcription on NVIDIA GPUs or Mac hardware and want to avoid cloud API costs or latency. It suits projects that process substantial audio volumes where speed matters. The README indicates the tool is opinionated about its defaults and recommends reviewing available CLI options to maximize throughput for your specific hardware. The project explicitly notes it works only on NVIDIA GPUs and Mac devices, so Windows users or those with other accelerators should verify compatibility before adoption.
The project evolved from a benchmarking showcase into a community-driven CLI tool. Development is responsive to community demand, with features added based on what users request. The tool maintains active support across different Whisper model variants, including the large-v3 and distil-whisper models, and provides installation flexibility through both standard package managers and pipx for isolated environments.
