Skip to content
← Models
Open weightsMeta (FAIR)/10 trustVerified

Meta Spirit LM

Open multimodal speech-text LM in Base and Expressive versions that keep the emotion in generated speech.

Visit official page →

About

Spirit LM is Meta FAIR's multimodal language model that treats speech and text as one token stream. Built on a 7B text model and continually trained on speech units, it comes in two versions: Base, which uses phonetic (HuBERT) units, and Expressive, which adds pitch and style units so it can detect and reproduce emotional states like anger or excitement in its speech output. The paper (arXiv 2402.05755) shows it learning ASR, TTS and speech classification few-shot, across modalities.

Meta released the weights, code and docs openly — but under the FAIR Noncommercial Research License, so this is research-only, not commercial open source. The paper appeared in February 2024 with a revised version that October, when Meta announced it publicly. All capability claims are author-reported; there is no independent benchmark.

Specs

Context window

Interleaved speech + text token stream; few-shot cross-modal tasks

Pricing

Free / open weights (research-only license)

Free quota

Fully open weights — no API, self-host only

API access

No official API — self-host the open weights (research, non-commercial)

Weights

Open weights

Released

Not disclosed

Benchmarks

Author-reported few-shot performance across ASR, TTS and speech classification; no independent verification

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-speechspeech-language-modelsresearchexpressive-ttsconversational-ai

Trust score — 7/10

/10 trustVerified

Scored 7/10 (TIER verified, metrics-v1). Evidence: published paper (arXiv 2402.05755, revised Oct 2024); Meta public announcement; open weights, code and docs released. Strengths: novel interleaved speech-text approach, expressive variant preserves emotion, open weights for research. Docked for: (1) FAIR Noncommercial Research License — research only, no commercial use; (2) all claims author-reported, no independent benchmark; (3) no maintained product or API. License: FAIR Noncommercial Research License (open weights, non-commercial).

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
8.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · Paper arXiv 2402.05755 (rev. Oct 2024); 7B text LM extended to speech via interleaved units
  • · Base (phonetic/HuBERT) + Expressive (pitch + style units) versions; few-shot ASR/TTS/classification
  • · Weights, code, docs open for research only; all capability claims author-reported
  • · Meta logo staged

Sources