#5151st of 76 in Model servingFor voice, speech and music
audio.cpp
C++ ggml runtime for local TTS, ASR, voice cloning and separation models
- Stars
- 4.3k
- License
- custom license
- Last commit
- Nov 2026
- Last release
- Nov 2026
- Language
- C++
297 on 27 Jun 20264,336 on 14 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
audio.cpp is a native C++ inference framework on ggml that runs 143 audio model families (250+ variants) for TTS, ASR, voice cloning, music generation, separation, VAD and diarization. It exposes them through a shared CLI, a server API, a C ABI and a built-in WebUI, with CUDA, HIP/ROCm, Vulkan, Metal, WebGPU, SYCL and CPU backends. Most models ship as GGUF packages, many with Q8 variants.
Who it is for: Self-hosters and developers running local speech models without Python stacks
Strengths
- Covers TTS, ASR, VAD, diarization, separation and music models under one CLI and server API
- Backends span CUDA, ROCm, Vulkan, Metal, WebGPU, SYCL and CPU, so no Python environments
- Built-in WebUI with Studio, Arena, voice library and TTS narration editor
- Q8 GGUF packages reported up to 1.53x faster and 37% lower peak VRAM than 16-bit
Weaknesses
- License is reported as NOASSERTION; model weights carry separate per-family licenses
- Fast-moving 0.9.x releases; HIP gfx1151_v1 kernels, avatar voice chat and LatentSync are experimental
- Model ports vary in maturity; some families are still under testing
- No Docker or compose files detected; builds are per-backend from source or portable builds
What it needs
- GPU optional
- Docker
- Needs ggml, GGUF model files (Hugging Face / ModelScope)
- Models: GGUF, Safetensors, Qwen3-TTS, Kokoro, Fish Audio S2
What changed
From the release notes. No flag means the note does not say so, not that the upgrade is safe.
Also in Model serving
See all 76| Rank | Project | Score |
|---|---|---|
| 1 | llama.cppC/C++ inference engine for GGUF LLMs and VLMs with an OpenAI-compatible server | 94 out of 100 |
| 2 | OllamaRuns open models locally through a CLI and REST API | 92 out of 100 |
| 3 | LocalAISelf-hosted model server with OpenAI, Anthropic and ElevenLabs-compatible APIs | 91 out of 100 |
| #4 | vLLMInference engine serving LLMs with PagedAttention and an OpenAI-compatible API | 88 out of 100 |
| #5 | strataRuns a 125B-parameter MoE model on one consumer GPU plus system RAM | 88 out of 100 |