transcribe.cpp: ggml-based cross-platform ASR library
Original: Transcribe.cpp
Why This Matters
Fills a critical gap in cross-platform local ASR inference with verified accuracy and broad GPU support.
Developer cjpais released transcribe.cpp v0.1.0, a ggml-based transcription library supporting 16 ASR model families (60+ models). It offers GPU acceleration via Vulkan, Metal, and CUDA, with numerically verified WER testing and bindings for Python, JavaScript/TypeScript, Rust, and ObjC/Swift.
transcribe.cpp is an open-source speech-to-text inference library built on ggml, created by the maintainer of the cross-platform desktop app Handy. The project was motivated by frustration with the existing ASR distribution landscape, which the author describes as limited to whisper.cpp and ONNX, with ONNX offering CPU-only performance and whisper.cpp lacking broad modern model support.
The library supports 16 ASR model families covering 60+ models, all published under the handy-computer Hugging Face organization. Every model has been numerically validated against its reference implementation and subjected to full Word Error Rate (WER) sweeps across thousands of utterances. Acceleration is supported via Vulkan, Metal, CUDA, and TinyBLAS, with benchmark results published for both a Ryzen 4750U (CPU + Vulkan on Fedora) and an Apple M4 Max.
The library supports both streaming and batch transcription, and is described as a near drop-in replacement for whisper.cpp. Official maintainer-supported bindings are available in Python, JavaScript/TypeScript, Rust, and ObjC/Swift, making it suitable for desktop and mobile application integration. The project is at v0.1.0 and the author is actively seeking bug reports from the community.