Transcribe.cpp: a ggml speech-to-text library with 60+ verified ASR models

Transcribe.cpp is a new open-source, ggml-based transcription library from the maintainer of the Handy speech-to-text app. It targets a gap the author kept hitting when shipping cross-platform local ASR: the practical choices were whisper.cpp or ONNX (with MLX bolted on for Apple), which meant maintaining multiple engines, porting models, and leaving GPU performance on the table. Other multi-model libraries existed but came with unknown provenance, unclear maintenance, and no benchmarks. The v0.1.0 release covers 16 ASR model families (60+ models), hardware acceleration across Vulkan, Metal, CUDA, and TinyBLAS, plus streaming and batch modes, and functions as a near drop-in replacement for whisper.cpp — including compatibility with existing .bin model files.

The project’s main differentiator is trust in inference quality. Every supported model is numerically validated against its reference implementation and put through full word-error-rate sweeps over thousands of utterances, with results published in the repo and on Hugging Face. It also ships maintainer-supported bindings for Python, JavaScript/TypeScript, Rust, and ObjC/Swift, treating real-world embeddability rather than demo code as a first-class goal. The author frames the effort as reducing the friction of distributing local inference so applications can run transcription on-device instead of sending audio to the cloud, citing faster-than-real-time performance on hardware as modest as an RK3566 board within a few watts.

The work was backed by Mozilla AI’s Builders-in-Residence program, with compute credits from Modal and CI from Blacksmith, and rests on the ggml ecosystem. As an early release it has acknowledged rough edges and some unsupported whisper.cpp flags, but its significance is in packaging validated, GPU-accelerated, broadly compatible local ASR into something an app developer can actually depend on and ship.