Skip to content

About

OpenAI-compatible speech & voice engine powered by F5-TTS, Faster-Whisper, and Edge-TTS

Resources

Contributing

Stars

7 stars

Watchers

0 watching

Forks

Repository files navigation

Nouverse Voice Compute 🎙️

FastAPI React CUDA OpenAI Whisper Qwen3--TTS License: MIT

Unified self-hosted voice intelligence stack and interactive dashboard. Built as a monorepo featuring Faster-Whisper Large-v3 for low-latency speech transcription/translation and Qwen3-TTS 1.7B Base for high-fidelity zero-shot voice cloning.


🌟 Key Features

  • 🗣️ State-of-the-Art Voice Cloning (Qwen3-TTS 1.7B Base): Autoregressive neural speech synthesis with 12Hz acoustic tokenizer. Clones custom timbres zero-shot from a short 3–10s audio reference.
  • 🎙️ Local STT & Speech Translation (Faster-Whisper Large-v3): CTranslate2 backend with Silero VAD, word-level timestamps, and native direct translation (task="translate").
  • 🌐 Realtime Indo → English Voice Translator: End-to-end pipeline: speak Indonesian, and the system translates the speech to English and speaks it back in your own cloned voice!
  • 💻 Interactive Studio Dashboard (apps/dashboard): Modern React 19 + DaisyUI web application served directly from the backend at http://localhost:8880. Includes Voice Clone Studio, Realtime Translator, and Whisper Playground.
  • 🔌 OpenAI Compatible: Standard drop-in endpoints for /v1/audio/transcriptions, /v1/audio/translations, and /v1/audio/speech.
  • 🧹 Zero VRAM Clutter: Only 2 active models maintained. Memory purge endpoint (/v1/models/unload) immediately frees VRAM.

🏗️ Monorepo Architecture

voice-compute/
├── apps/
│   ├── voice-server/            # FastAPI Python backend (Port 8880)
│   │   ├── core/                # Configuration & device management
│   │   ├── engines/             # Faster-Whisper & Qwen3-TTS wrappers
│   │   ├── routes/              # Audio, voice cloning & pipeline endpoints
│   │   ├── main.py              # App entrypoint (serves API & Dashboard)
│   │   └── requirements.txt
│   │
│   └── dashboard/               # Interactive React 19 + Vite web studio
│       ├── src/
│       │   ├── components/      # VoiceCloneTab, TranslatorTab, TranscribeTab
│       │   ├── App.tsx          # Main studio interface & telemetry
│       │   └── main.tsx
│       ├── package.json
│       └── vite.config.ts
│
├── models/                      # Exactly 2 active models (Git-ignored)
│   ├── faster-whisper-large-v3/ # CTranslate2 model weights
│   └── qwen3-tts/               # Qwen3-TTS-12Hz-1.7B-Base + Tokenizer
│
├── voices/                      # Reference audio clips (.wav & .txt transcripts)
├── scripts/
│   └── download_models.py       # Automated model downloader
├── start.bat                    # Windows one-click start script
├── start.sh                     # Linux/macOS start script
└── requirements.txt

🚀 Quick Start

1. Launch with One Click

# On Windows:
start.bat

# On Linux / macOS:
chmod +x start.sh
./start.sh

Open your browser at: 👉 http://localhost:8880 to access the Voice Compute Studio Dashboard!


2. Manual Development Setup

Backend (apps/voice-server):

python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install Python requirements
pip install -r requirements.txt

# Run server
export PYTHONPATH="$(pwd)/apps/voice-server"
python -m uvicorn main:app --app-dir apps/voice-server --host 0.0.0.0 --port 8880 --reload

Frontend Dashboard (apps/dashboard):

cd apps/dashboard
bun install
bun run dev     # Starts Vite dev server on http://localhost:5173
bun run build   # Compiles production bundle into dist/

📡 API Endpoints

Method Endpoint Description
POST /v1/audio/transcriptions OpenAI-compatible Whisper transcription
POST /v1/audio/translations Whisper direct foreign-speech to English translation
POST /v1/audio/speech Qwen3-TTS zero-shot voice cloning synthesis
POST /v1/audio/pipeline/translate-voice End-to-end Indonesian speech $\rightarrow$ Cloned English voice
GET /v1/audio/voices List installed voice profiles
POST /v1/voices/clone Upload and register a new voice reference
POST /v1/models/unload Purge AI models from GPU VRAM
GET /status GPU device & VRAM usage telemetry

📄 License

Distributed under the MIT License. Copyright © 2026 Nouverse Technologies.

About

OpenAI-compatible speech & voice engine powered by F5-TTS, Faster-Whisper, and Edge-TTS

Resources

Contributing

Stars

7 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages