Sonus — Speak freely

Multi-engine, offline audio hub (Speech, Music, SFX, STT) on your Mac. No cloud. No API keys. No one listening.

Runs on MLX and ONNX Runtime — Apple Silicon for MLX engines, any platform for ONNX engines (Kokoro, Piper).

About the name: Sonus is Latin for "sound" (/ˈsoː.nus/). Felt right for a project about making machines talk.

ℹ️ Not affiliated with any company, service, or organization named Sonus. Just a coincidence. We're an independent open-source thing.

Opencode — vibecoded by AI, tested by humans.

---

What is this?

Sonus turns text into speech/music/SFX (and speech into text) using whatever engine you throw at it. Multiple engines, one unified API. Run it locally, hit the endpoint, get audio back. Zero data leaves your machine.

Think of it as a local audio hub — TTS via Qwen3, Kokoro, Piper, Chatterbox, CosyVoice, and Fish Speech; Speech-to-Music/SFX via MusicGen, Riffusion, and Stable Audio Open; and STT via Whisper. All offline, all local.

---

Run in the cloud

No Mac? No problem. Kokoro and Piper (ONNX engines) work on any platform. Click a badge to open a pre-configured notebook — it installs everything, downloads models, starts the server, and gives you a public URL.

---

Engines at a glance

More on each engine:

• Qwen3 — custom voice, voice design, voice cloning
• Kokoro — 54 built-in voices, 9 languages
• Piper — fastest inference, widest language support
• Chatterbox Turbo — voice cloning via MLX
• CosyVoice2 — zero-shot TTS and prosody control
• Fish Speech — multilingual TTS with voice cloning
• MusicGen / AudioGen — text-to-music and text-to-sfx
• Riffusion — text-to-music via diffusion
• Stable Audio Open — stereo music and SFX generation
• Whisper MLX — speech-to-text via MLX
• API Reference — full endpoint docs

---

What you'll need

Local (macOS Apple Silicon):

• A Mac with Apple Silicon (M1, M2, M3, M4 — anything with Metal)
• Python 3.13+ (brew install python@3.13)
• ffmpeg — brew install ffmpeg

Running in Docker? Only Kokoro and Piper work there. Qwen3 and Chatterbox need the Metal GPU backend. See each engine's doc for details.

---

Get started in 30 seconds

git clone https://github.com/YOUR_USERNAME/sonus.git
cd sonus

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
brew install ffmpeg

Download models for at least one engine (check the engine docs above), then:

source venv/bin/activate
python server.py
# Listening on http://0.0.0.0:8000

Interactive API docs: http://localhost:8000/api-docs

---

Try it

curl -X POST http://localhost:8000/tts \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world", "model": "kokoro-v1.0", "speaker_name": "af_heart"}' \
  --output hello.mp3

Speaking OpenAI's language

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "kokoro", "input": "Hello world", "voice": "af_bella"}' \
  --output hello.mp3

Drop-in replacement for POST /v1/audio/speech. Your existing OpenAI TTS code works without changes — just point it at http://localhost:8000.

Speech to text (OpenAI-compatible)

curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -F "file=@speech.mp3" \
  -F "model=whisper-base" \
  -F "language=en" \
  -F "temperature=0.0"

{"text": "Hello world, this was transcribed locally."}

---

Documentation

---

How it's built

server.py               # FastAPI — the brain
src/
  engines/
    base.py             # The contract every TTS/generation engine signs
    qwen.py             # Qwen3 (MLX)
    chatterbox.py       # Chatterbox Turbo (MLX)
    kokoro.py           # Kokoro (ONNX)
    piper.py            # Piper (ONNX)
    cosyvoice.py        # CosyVoice2 (PyTorch)
    fish_speech.py      # Fish Speech (PyTorch/MLX)
    musicgen.py         # MusicGen & AudioGen (MLX)
    riffusion.py        # Riffusion (PyTorch/Diffusers)
    stable_audio.py     # Stable Audio Open (PyTorch/Diffusers)
  stt/
    base.py             # The contract every STT engine signs
    whisper_mlx.py      # Whisper via MLX (Apple Silicon)
static/                 # Web UI — Vue 3 (CDN), no build step
docs/                   # API, engine, development, and MCP docs
models/                 # Downloaded models (gitignored)
voices/                 # WAVs for voice cloning
sfx/                    # Sound effects database
outputs/                # Generated audio files

---

Built with

---

Disclaimer

Yes, this thing can make audio that sounds like real people. You're responsible for what you do with it.

• Don't impersonate people without their consent.
• Don't create deceptive, fraudulent, or misleading content.
• Do respect the laws where you live.
• The authors assume zero liability for misuse.

Use it wisely. Or don't — but that's on you.

---

License

MIT — do what you want with it, just keep the notice.