handy-computer / transcribe.cpp
A unified C++ inference library bringing 16+ speech-to-text model families to the ggml/GGUF ecosystem with multi-backend GPU acceleration.
星标趋势
AI 分析
项目摘要
transcribe.cpp is a C/C++ speech-to-text inference library that runs 16+ STT model families (60+ variants) via GGUF models on the ggml runtime, with Metal, Vulkan, and CUDA GPU backends plus a tinyBLAS-accelerated CPU path. It supports both streaming and batch transcription, with every published model numerically verified and WER-tested against reference implementations.
为什么值得关注
It unifies an exceptionally broad range of ASR model families — from Whisper and Canary to Qwen3-ASR, Voxtral, and Parakeet — under a single inference library with multi-backend GPU acceleration, filling a gap in the ggml ecosystem for production-grade speech-to-text beyond just Whisper.
优势
- Unmatched model family breadth: 16+ ASR architectures and 60+ variants in one library
- Multi-backend GPU acceleration (Metal, Vulkan, CUDA) plus optimized CPU path
- Numerically verified and WER-tested models against reference implementations
- Active development with 12 releases in 6 months and comprehensive per-model documentation
局限性
- Single contributor project, creating bus-factor risk
- No Docker support for containerized deployment
- Relatively new project (under 6 months old) with limited production track record
使用场景
- Embedding diverse ASR capabilities into C/C++ applications with a single dependency
- Edge and on-device speech-to-text deployment with GPU acceleration
- Benchmarking and comparing multiple ASR model families under identical inference conditions
- Streaming transcription for real-time applications