0xshug0/audio.cpp

audio.cpp is a C++ runtime for running audio AI models locally without any Python installation. If you have ever wanted to use text-to-speech, voice cloning, speech recognition, or music generation on your own computer but found that setting up the Python packages for these tools is complicated and fragile, audio.cpp offers an alternative: a single compiled program that handles many different audio model families through a common interface. The framework is built on top of ggml, the same low-level computation library used by tools like llama.cpp for running large language models locally. This means audio.cpp can run audio models efficiently on CUDA graphics cards, with reported speed improvements of 1.8 to 5 times faster than the equivalent Python implementations for some models. For example, a text-to-speech model called VibeVoice can generate about 94 minutes of speech in roughly 18 minutes on a compatible GPU. The list of supported model types is broad. On the text-to-speech side, there are over 20 different model families available, supporting dozens of languages including English, Chinese, Japanese, German, French, and many others. Voice cloning, which copies a voice from a short audio sample, is supported in several of these. On the audio understanding side, there are speech recognition models, voice activity detection (which detects when someone is speaking versus silence), and speaker diarization (which identifies which speaker said which part). Music generation and audio source separation, which splits a mixed audio track into its parts such as vocals and instruments, are also included. The tool runs from the command line and also includes an API server mode for integration with other software. There is experimental support for defining multi-step audio processing workflows in a configuration file, so you can chain operations like transcription followed by voice conversion without writing custom code. Building from source requires a C++ compiler and CMake. CUDA support requires the CUDA toolkit. The project is under active development. The full README is longer than what was shown.