audio.cpp runs AI audio generation locally on your computer. It supports text-to-speech, voice cloning, sound effects, music generation, and audio-conditioned music generation. Select a model for your task, download it from the track settings, then generate normally.
Quick start
- Add an audio.cpp audio track.
- Select a model based on the job you want to do.
- Click Download selected model.
- Enter text, a prompt, lyrics, or music tags as appropriate.
- Generate audio. The first run can take longer while files are prepared.
Tip: Q8 models are the recommended practical choice for most systems. F16/BF16 versions use more memory and are intended for stronger GPUs.
Speech and voice models
| Model | Best use case |
|---|---|
| Supertonic 3 Q8 | Fast multilingual text-to-speech on low-end GPUs or CPU. |
| MOSS-TTS Nano 100M Q8 | Small, fast speech model with optional voice cloning. |
| PocketTTS English / German / Italian / Portuguese / Spanish Q8 | Fast low-resource speech with built-in voices and optional cloning. |
| Irodori TTS 500M Q8 | Japanese speech and voice cloning. |
| Irodori Voice Design 600M Q8 | Japanese speech with a described or designed voice. |
| OmniVoice Q8 | Broad language support, voice cloning, and voice design. |
| Chatterbox Q8 | Expressive voice cloning. Requires a reference voice clip. |
| MioTTS 1.7B Q8 | Higher-quality voice cloning for mid-range GPUs. |
| Qwen3 TTS Base 1.7B Q8 | Versatile multilingual voice cloning. |
| Qwen3 TTS CustomVoice 1.7B Q8 | High-quality speech with built-in voice identities. |
| Qwen3 TTS VoiceDesign 1.7B Q8 | Create speech from a description of the desired voice. |
| IndexTTS2 Q8 | Expressive English and Chinese voice cloning. |
| VibeVoice 1.5B Q8 | Long-form dialogue and extended speech. |
| VoxCPM2 Q8 | High-quality controllable speech and voice cloning. |
| MOSS-TTS Local 1.5 Q8 | Higher-quality multilingual voice cloning for strong GPUs. |
| Fish Audio S2 Pro Q8 / BF16 | Premium expressive speech and voice cloning. BF16 needs substantially more GPU memory. |
| Higgs Audio v3 TTS 4B Q8 / BF16 | Premium voice-clone quality. Requires a reference voice clip and a strong GPU. |
| Vevo2 Q8 / F16 | Speech or singing from a reference voice clip. |
Music and sound-effect models
| Model | Best use case |
|---|---|
| Stable Audio 3 Small Music Q8 / F16 | Music generation on modest hardware. Supports source-audio conditioning and inpainting. |
| Stable Audio 3 Small SFX Q8 / F16 | Sound effects, ambience, and short audio prompts. Supports source-audio conditioning and inpainting. |
| Stable Audio 3 Medium Q8 / F16 | Higher-quality music generation for stronger GPUs. |
| HeartMuLa Q8 / F16 | Song generation from lyrics and music tags, including multilingual music. |
Useful settings
- Backend: Use best unless you specifically need CUDA, Vulkan, or CPU mode.
- Voice reference: Add a clean speaker clip when using a voice-cloning model. Audio is converted automatically for compatibility.
- Source audio: Available for Stable Audio models. Use it for audio variation or inpainting.
- Music tags: For HeartMuLa, describe genre, mood, instruments, tempo, and vocals.
- Seed: Set a track seed to reuse it across items. Leave both track and item seed at 0 to create and save a random item seed.
Model quality, generation speed, and memory use vary significantly. Start with a Q8 model and move to F16/BF16 only when you have sufficient GPU memory.