Skip to content

SPARKTTS.md

Latest commit

 

History

History
135 lines (92 loc) · 3.17 KB

SPARKTTS.md

File metadata and controls

135 lines (92 loc) · 3.17 KB

Spark-TTS - Optional Local Voice Cloning

Spark-TTS is an optional LLM-based TTS model (0.5B parameters) that provides zero-shot voice cloning using your character .wav files.

Why Spark-TTS?

  • ✅ Zero-shot voice cloning - uses character voice files
  • ✅ Fully local - no API keys needed
  • ✅ GPU accelerated - fast inference with CUDA
  • ✅ Cross-platform - Windows, Linux, macOS

Requirements

  • Python: 3.11+ (recommended for CUDA support)
  • GPU: CUDA-capable GPU (optional, will use CPU)
  • Disk: ~5GB for model files
  • RAM: 8GB+ (16GB+ recommended)

Quick Setup

Automated Setup (Recommended)

# GPU (CUDA) version
python setup_sparktts.py

# CPU-only version
python setup_sparktts.py --cpu-only

Manual Setup

  1. Install dependencies:

    # For GPU (CUDA 12.4)
pip install -r requirements_cuda_sparktts.txt

# For CPU only
pip install -r requirements_cpu_sparktts.txt

  2. Download model (~4GB):

    python download_sparktts_model.py

  3. Configure .env:

    TTS_PROVIDER=sparktts
SPARKTTS_MODEL_DIR=pretrained_models/Spark-TTS-0.5B
SPARKTTS_MAX_CHARS=1000

  4. Run the app:

    uvicorn app.main:app --host 0.0.0.0 --port 8000

Python Version Notes

  • Python 3.11+: Recommended - full PyTorch 2.6+ CUDA support

Performance

Hardware Speed Quality
CPU Slow (~30-60s) Good
GPU (CUDA) Fast (~2-5s) Good

Troubleshooting

"TTS Provider is blank in UI"

  • Ensure TTS_PROVIDER=sparktts in .env
  • Restart the application

"CUDA not available"

# Verify CUDA installation
python -c "import torch; print(torch.cuda.is_available())"

# Reinstall CUDA PyTorch
pip install torch torchaudio torchvision --index-url https://download.pytorch.org/whl/cu124

"numpy version incompatible"

pip install "numpy>=1.21.6,<1.28.0"

Model download fails

  • Check internet connection
  • Ensure ~5GB free disk space
  • Try manual download from Hugging Face

Alternative TTS Options

Spark-TTS is optional. The app also supports:

  • OpenAI TTS - Cloud-based, fast, high quality
  • ElevenLabs - Cloud-based, excellent quality
  • Kokoro TTS - Local, CPU-friendly, very fast

Files Overview

voice-chat-ai/
├── sparktts/              # Spark-TTS core modules (integrated)
├── cli/                   # SparkTTS class and inference
├── download_sparktts_model.py  # Model download script
├── setup_sparktts.py      # Automated setup script
├── requirements_cpu_sparktts.txt   # CPU dependencies
├── requirements_cuda_sparktts.txt  # CUDA dependencies
└── pretrained_models/     # Model storage (gitignored)

Uninstalling

To remove Spark-TTS:

  1. Delete pretrained_models/ folder
  2. Delete sparktts/ and cli/ folders
  3. Set TTS_PROVIDER to another option in .env
  4. Remove Spark-TTS packages: pip uninstall einx einops omegaconf hydra-core soxr

Note: This is an experimental feature. For production use, consider OpenAI or ElevenLabs TTS.