Spark-TTS: Synthesize Speech with Qwen2.5

Overview

Spark-TTS is an inference system built entirely on Qwen2.5 for generating natural-sounding speech from text. Rather than using a separate generation model for acoustic features, it reconstructs audio from speech codes predicted by the LLM, aiming to simplify the synthesis pipeline.

It is designed for Chinese and English speech, including zero-shot voice cloning across languages and code-switching. The repository provides PyTorch inference code and a web UI, with Linux instructions and a separate Windows installation guide.

Key Features

• Zero-shot voice cloning from reference audio and its transcript.
• Chinese and English speech synthesis, including cross-lingual and code-switching scenarios.
• Voice creation with controls for characteristics such as gender, pitch, and speaking rate.
• Command-line inference and a Gradio-based web UI for cloning and voice creation.
• Optional Nvidia Triton and TensorRT-LLM inference serving reference.
• Model weights are available separately from the code through Hugging Face.

Use Cases

• Researchers exploring LLM-based speech generation can run the released inference code and consult the linked paper.
• Developers building Chinese-English speech prototypes can test voice cloning and code-switching with reference audio.
• Speech tool builders can evaluate adjustable synthetic voices for applications such as personalized speech or assistive technology.
• Teams deploying inference on Nvidia hardware can use the Triton and TensorRT-LLM reference as a starting point.

Getting Started

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

Download the SparkAudio/Spark-TTS-0.5B model from Hugging Face before running inference. See the README for model setup, CLI and web UI instructions, and runtime deployment details.

Alternatives

• GPT-SoVITS: GPT-SoVITS provides a WebUI for training and inference, including fine-tuning on a small speaker dataset.
• VoxCPM: VoxCPM uses a tokenizer-free architecture and targets multilingual speech, voice design, and cloning with a 2B-parameter model.
• ChatTTS: ChatTTS focuses on dialogue-style Chinese and English speech, with multiple speakers and controls for pauses and laughter.
• MOSS-TTS: MOSS-TTS covers long-form narration, sound effects, voice design, and streaming across multiple model architectures.

Considerations

• The repository is inactive by the supplied health data, with no recent commits, releases, merged pull requests, or closed issues. Five contributors are recorded, and the data does not flag contributor concentration as a project risk.
• The documented setup targets Linux, Python 3.12, and PyTorch 2.5 or later. Windows users are directed to an issue for installation guidance.
• The model weights are a separate download. Inference and the Triton serving reference use GPU-oriented device settings, so check hardware and deployment requirements before adopting it.
• Training code and the VoxBox dataset are not included according to the project’s to-do list.