terrancemoon9/Boovore_Multi-Engine-TTS-Studio

🎙 Boovore — Multi-Engine TTS Studio

Boovore is a self-hosted, GPU-accelerated Text-to-Speech studio with 6 best-in-class engines and a built-in audiobook generator. Run it on any CUDA machine (tested on RTX 3090) via a clean, dark-mode web UI.

Name: Boovore = Book + Devour — built to devour books in audio.

[image]

✨ Engines

Voxtral uses vLLM-Omni (mistralai/Voxtral-4B-TTS-2603) with voice cloning via a reference WAV. Start it separately with python3 voxtral_server.py.

⚙️ CPU / GPU — HuggingFace Space Settings

In your Space → Settings → Variables and secrets, set:

Default is all — on free CPU tier, set kokoro,f5 to avoid crashes.

For Voxtral, also set VOXTRAL_URL to point to your vLLM-Omni server (default: http://localhost:8000).

🚀 Quick Start (Vast.ai / GPU server)

1. Install dependencies

# PyTorch nightly CUDA 12.8 (required)
pip3 install --pre torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

# Register torch libs so torchaudio can find libtorch
echo "/usr/local/lib/python3.12/dist-packages/torch/lib" > /etc/ld.so.conf.d/torch.conf && ldconfig

# Core engines
pip3 install faster-qwen3-tts kokoro f5-tts fastapi uvicorn[standard] python-multipart

# Chatterbox (Python 3.12 fix)
pip3 install conformer==0.3.2 --no-build-isolation
git clone https://github.com/resemble-ai/chatterbox /tmp/chatterbox
cd /tmp/chatterbox && pip3 install -e . --no-deps && cd /root

# Fish-Speech 1.5
git clone https://github.com/fishaudio/fish-speech /tmp/fish-speech
cd /tmp/fish-speech && git checkout v1.5.1
pip3 install -e . --no-deps
huggingface-cli download fishaudio/fish-speech-1.5 --local-dir /root/fish-speech-model

2. (Optional) Start Voxtral TTS server

Voxtral requires a separate vLLM-Omni process (~8 GB VRAM). Needs a HuggingFace token — accept the CC BY-NC license at mistralai/Voxtral-4B-TTS-2603 first.

pip install "vllm[audio]>=0.18.0" httpx soundfile
export HF_TOKEN=hf_xxxx
nohup python3 voxtral_server.py >> /root/voxtral.log 2>&1 &
# Wait 5-10 min for model download + load (first run only)

Optionally generate a narrator reference WAV (for voice cloning):

# While the Qwen3 server is running:
python3 make_narrator_reference.py
# Output: /workspace/narrator_reference.wav

3. Start the main server

nohup python3 server.py --port 7860 >> /root/server.log 2>&1 &

3. Open the UI

# Local SSH tunnel
ssh -p <PORT> root@<HOST> -L 7860:localhost:7860 -N
# Then open http://localhost:7860

📖 Features

• —TTS Studio — one-click engine selector (8 pills), single generate button
• —Audiobook Generator — import .txt / .pdf / .epub, auto-detect chapters, batch generate with any engine, download per chapter or merge into one WAV
• —Voice Cloning — upload a reference audio clip (Chatterbox, F5-TTS, Fish-Speech, Qwen3)
• —Real-time metrics — TTFA, RTF, duration, buffer
• —Light / dark theme
• —Streaming audio (Qwen3) with CUDA Graphs

🗂 Project Structure

server.py                    — FastAPI backend (6 engines)
index.html                   — UI single-page (vanilla JS, no frontend deps)
voxtral_server.py            — vLLM-Omni server manager (start/stop/status)
make_narrator_reference.py   — Generate narrator reference WAV via Qwen3
narrator_reference.wav       — (generated) voice clone reference for Voxtral
requirements.txt
Dockerfile

⚙️ Requirements

• —Python 3.12+
• —CUDA 12.8 (RTX 3090 or better recommended)
• —PyTorch nightly cu128 (2.12.0.dev+)
• —VRAM: 8 GB minimum, 24 GB to run all engines simultaneously

📦 Models (auto-downloaded)

🏷️ GitHub Topics

text-to-speech tts voice-cloning audiobook french-tts kokoro f5-tts fish-speech chatterbox qwen3 voxtral mistral vllm fastapi cuda self-hosted gpu french multilingual

Credits

• —faster-qwen3-tts — Qwen3-TTS engine
• —Fish-Speech — fishaudio
• —Chatterbox — ResembleAI
• —F5-TTS — SWivid
• —Kokoro — hexgrad
• —Voxtral — Mistral AI (mistralai/Voxtral-4B-TTS-2603, CC BY-NC)
• —French prosody preprocessing inspired by arXiv:2508.17494

MIT License