Skip to content

Spark-TTS

An LLM-based TTS model with single-stream decoupled tokens

Visit official site

What Spark-TTS is

The model is described as an efficient LLM-based text-to-speech system with single-stream decoupled speech tokens that rebuilds audio directly from predicted codes.

What you can do with it

  • Synthesise speech from text with a large language model
  • Run zero-shot text-to-speech synthesis across languages
  • Create virtual voices by adjusting parameters
  • Control speech rate, pitch and gender

Application scenarios

  • Researchers run speech synthesis experiments and studies
  • Teachers generate speech examples in different languages and styles
  • Companies build personalised voice prompts or voice navigation

Getting started

  • Clone the project repository
  • Create and activate a Conda environment with Python 3.12
  • Install the listed dependencies
  • Download pretrained models from Hugging Face or with git lfs

Pros & cons

✓ What we like

  • Zero-shot synthesis works across languages and code switching
  • High-quality Chinese and English synthesis is listed

! What to watch out for

  • The model is described as aimed at academic research and lawful applications
  • Setup requires cloning a repository and downloading pretrained models

FAQ

Which languages are named?

The feature list names Chinese and English for high-quality synthesis.

What can be adjusted in a virtual voice?

The material lists speech rate, pitch and gender as adjustable parameters.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review