Skip to content

Llasa-3B

A LLaMA-based TTS model for Chinese and English speech

Visit official site

What Llasa-3B is

The model is described as a text-to-speech system built on the LLaMA architecture for Chinese and English speech synthesis. It combines XCodec2 speech coding to convert text into natural, fluent speech.

What you can do with it

  • Convert Chinese and English text into speech efficiently
  • Generate more natural speech from a given voice prompt
  • Use custom training to fit different voice styles
  • Explore and extend the open-source model

Application scenarios

  • Generate Chinese and English speech for audiobook platforms
  • Build multilingual voice assistants with fluent interaction
  • Produce course narration for online education platforms

Getting started

  • Install XCodec2 and the related dependency libraries
  • Load AutoTokenizer and AutoModelForCausalLM from Hugging Face
  • Format the input text and generate speech codes
  • Decode the speech codes into a speech waveform

Pros & cons

✓ What we like

  • Based on the LLaMA architecture with language understanding
  • Open-source design lets developers explore and extend it

! What to watch out for

  • Speech synthesis centres on Chinese and English
  • Generation runs through speech codes decoded into waveforms

FAQ

Which languages does it synthesise?

The material names Chinese and English speech synthesis.

Is the model open?

The material describes it as open-source for developers to explore and extend.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review