handy-computer/transcribe.cpp

transcribe.cpp is a C and C++ library for turning speech audio into text. It runs on the ggml engine, the same style of runtime used by projects like whisper.cpp, and supports GPU backends including Metal on Apple Silicon, Vulkan, and CUDA, plus a fast CPU only path when no GPU is available. The standout feature is breadth: the library supports 16 different families of speech recognition models and more than 60 variants in total, covering both streaming audio and batch processing of full recordings. Supported families include well known names like OpenAI's Whisper and NVIDIA's Parakeet and Canary, along with a range of newer or more specialized models such as Moonshine, GigaAM, Qwen3-ASR, and a medical dictation model called MedASR. Every model published by the project is checked against its original reference implementation to confirm the transcriptions match. Getting started means building the project with CMake, which automatically turns on Metal support on Apple hardware, or can be configured for Vulkan or CUDA on Linux and Windows. Pre-built model files are hosted on Hugging Face and linked from documentation for each model family, so most users can download a ready to use model rather than converting one from scratch. For those who do need to convert a model, the project includes a script that pulls from NVIDIA's NeMo checkpoints. A separate quantization tool can shrink models to smaller file sizes at a small cost to precision. Once built, transcription happens through a command line tool that takes a 16kHz mono WAV audio file and a chosen model file as input. Other audio formats need to be converted first using a tool like ffmpeg. For developers who want to use the library from another programming language, official bindings are provided for Python, TypeScript and JavaScript, Rust, and Swift or Objective-C. This project fits developers building applications that need offline or self-hosted speech-to-text, such as transcription tools, voice assistants, or accessibility features, where running many different model options through one consistent interface is valuable.