Spark-TTS is an LLM-based text-to-speech system designed for efficient, high-quality voice synthesis and zero-shot voice cloning. It targets researchers and developers needing natural, controllable speech generation across languages, offering a streamlined approach by directly generating audio from LLM-predicted speech tokens.

How It Works

Spark-TTS leverages the Qwen2.5 LLM to directly reconstruct audio from predicted speech tokens, eliminating the need for separate acoustic models. This single-stream, decoupled token approach simplifies the pipeline, enhances efficiency, and enables high-fidelity voice cloning and controllable speech synthesis (e.g., adjusting pitch, speaking rate) without explicit fine-tuning for each voice.

Quick Start & Requirements

• Install: Clone the repository and install dependencies via Conda:

  git clone https://github.com/SparkAudio/Spark-TTS.git
  cd Spark-TTS
  conda create -n sparktts -y python=3.12
  conda activate sparktts
  pip install -r requirements.txt
  

• Model Download: Use Hugging Face snapshot_download or git clone with git-lfs.
• Basic Usage: Run inference via bash example/infer.sh or python -m cli.inference.
• Web UI: Start with python webui.py --device 0.
• Prerequisites: Python 3.12, Conda, git-lfs. GPU recommended for inference.
• Docs: Runtime section for Triton/TensorRT-LLM deployment.

Highlighted Details

• Supports zero-shot voice cloning for cross-lingual and code-switching scenarios.
• Enables controllable speech generation by adjusting parameters like gender, pitch, and speaking rate.
• Offers Nvidia Triton Inference Serving support with benchmark results showing low latency and high RTF (e.g., 0.1362 RTF on L20 GPU).
• Bilingual support for Chinese and English.

Maintenance & Community

• Paper published on arXiv: Spark-TTS.
• Nvidia Triton Inference Serving support added March 12, 2025.
• Community contributions noted for Windows installation guide.

Licensing & Compatibility

• License: Not explicitly stated in the README.
• Compatibility: Intended for academic research, educational purposes, and legitimate applications. Usage disclaimer strongly advises against unauthorized voice cloning, impersonation, fraud, scams, deepfakes, or illegal activities.

Limitations & Caveats

The project is primarily focused on inference and provides a usage disclaimer emphasizing responsible use. Training code and dataset (VoxBox) are listed as future releases (To-Do).