audio.cpp

audio.cpp runs audio models locally for developers building speech tools and people generating or processing audio on their own hardware. It brings multiple audio tasks into one C++ runtime built on ggml.

• Speech generation and voices: Generate speech with Kokoro or Qwen3-TTS, clone voices with Chatterbox, and hold speech-to-speech conversations with LFM2.5-Audio.
• Transcription and speech analysis: Turn speech into text with Qwen3-ASR or Nemotron ASR, identify speaker turns, detect speech activity, and align text with audio using supported models.
• Music and sound generation: Create music with ACE-Step or MiniMax Music 3, generate songs conditioned on lyrics with YuE2, and produce sound effects with Stable Audio.
• Audio cleanup and conversion: Remove noise, separate vocals or speakers, restore recordings, and change voices with models such as HTDemucs, AudioSR and SeedVC.
• Model comparison: The browser-based Arena lets you compare speech generation, voice conversion and transcription models using the same input, then review outputs and metrics.

It runs on Windows, Linux and macOS through a command-line interface, server API or WebUI, and also powers audio tasks in Unsloth Studio. It supports CPU-only machines, NVIDIA and AMD GPUs, and Apple Silicon through CUDA, HIP/ROCm, Vulkan and Metal backends. GGUF support includes quantized packages, while some models accept Safetensors. Streaming and language coverage vary by model. Model weights have separate licenses that determine commercial-use permissions; model-based audio utilities need separately downloaded weights.