The developer behind the Handy transcription application has introduced transcribe.cpp, a C and C++ library intended to make local automatic speech recognition easier to ship across desktop and mobile platforms. The initial v0.1.0 release uses ggml and is explicitly described as early software with rough edges that need community testing.
The project grew from the maintainer’s difficulty distributing Handy with a fragmented inference stack. Existing choices included whisper.cpp and ONNX, while Apple’s MLX could add another engine to maintain. The new library aims to combine broad model support, hardware acceleration and a compact deployment footprint suitable for applications rather than demonstrations. Its stated targets include macOS, Windows and Linux.
Accuracy validation is central to the release. According to the project author, every model published through the handy-computer organization on Hugging Face is checked numerically against its reference implementation. The team also runs word-error-rate sweeps over thousands of utterances and publishes results with the repository and model files. Those are developer-reported tests, but they address a practical concern: converted models may run successfully while producing results that differ from the original implementation.
The library includes benchmarks from a Ryzen 4750U system using CPU and Vulkan on Fedora, as well as an Apple M4 Max. Vulkan support was a priority because the maintainer regards GPU inference as a baseline requirement for practical local applications. The author also reports that an RK3566 system can transcribe faster than real time on its CPU, illustrating the project’s focus on low-power devices.
Compatibility with whisper.cpp influenced the interface. transcribe.cpp can run widely used `.bin` model files and is intended to cover most common replacement scenarios, although some whisper.cpp flags and features remain unsupported. First-party language bindings are also part of the design; Rust was an immediate requirement for Handy, and the maintainer selected four languages for initial support while inviting contributors to maintain others.
The project received support from Mozilla AI’s Builders incubator program, ggml contributors and Modal credits used for error-rate testing. Its broader proposition is privacy and portability: accurate transcription can occur on a user’s own hardware without routinely sending voice recordings to a cloud service.
