Skip to content

Spirit LM

A language model mixing text and speech

Visit official site
Category AI models

What Spirit LM is

The model is described as a foundational multimodal language model that mixes text and speech in a single token stream, trained by word-level interleaving.

What you can do with it

  • Feed text and speech into one sequence
  • Use phoneme units in the Base version
  • Add pitch and style units in the Expressive version

Getting started

  • Read the GitHub page or the paper
  • Choose a version and download its weights
  • Prepare a parallel speech-text corpus

Pros & cons

✓ What we like

  • Text semantics and speech expression are both stated
  • New tasks can be learned from few examples

! What to watch out for

  • Two versions are described with different units
  • The stated base is a 7B text model

FAQ

What are the two versions?

The material names a Base version with phoneme units and an Expressive version with pitch and style units.

Which tasks are named?

ASR, TTS and speech classification are named as cross-modal tasks.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review