Transcribe.cpp
Points and comments are a snapshot, not live.
A new ggml-based library supports 16 ASR model families with GPU acceleration.
transcribe.cpp is a ggml-based transcription library supporting 16 ASR families (60+ models) with acceleration via Vulkan, Metal, CUDA, and TinyBLAS. Every model is numerically validated and WER-tested against reference implementations. It supports streaming and batch transcription, serves as a drop-in replacement for whisper.cpp, and ships with bindings for Python, JavaScript/TypeScript, Rust, and ObjC/Swift. The author created it to solve cross-platform distribution challenges for Handy, his speech-to-text app. Mozilla AI, Modal, and Blacksmith supported development.
What commenters are saying
Commenters broadly praised the project's approach to local ASR and its practical usefulness. The author confirmed diarization is in progress, debuting this week. Several users reported positive experiences with Handy using the new streaming models like Parakeet and Nemotron. A debate emerged about whether using AI for transcription shapes one's thinking over time, with some pushing back that transcription models aren't LLMs. One user noted the Rust bindings make it a strong candidate for Tauri-based apps.