2 days ago
- transcribe.cpp is a ggml-based open-source transcription library supporting various ASR models with numerical validation and WER testing.
- It addresses challenges in distributing cross-platform speech-to-text applications, offering GPU acceleration via Vulkan, Metal, CUDA, and TinyBLAS.
- The library supports over 16 ASR families (60+ models), streaming/batch transcription, and is a near drop-in replacement for whisper.cpp.
- It includes maintainer-supported bindings for Python, JavaScript/TypeScript, Rust, and ObjC/Swift to facilitate wide distribution.
- Motivation stems from the need for a reliable, embeddable library that ensures inference accuracy and performance across Mac, Windows, and Linux.
- Key features include numerical verification against reference implementations, benchmark results for CPU/Vulkan, and open-source contributions.
- The project aims to make local speech-to-text more accessible, enabling real-time transcription on devices like RK3566 without cloud dependency.
- Acknowledgments to Mozilla AI, ggml contributors, Modal, Blacksmith, and Hugging Face for their support in development and testing.