Model
Inference & Serving
Sesame CSM
1Conversational 2speech 3model 4for 5natural 6dialogue
the six Ws · specification
W1
Who
Sesame AI Labs released the Conversational Speech Model, CSM.
W2
What
CSM is a speech generation model built on a Llama backbone that produces contextually natural, conversational sounding speech from text and prior audio context.
W3
Where
Hosted on GitHub at SesameAILabs/csm and on Hugging Face under sesame.
W4
When
Released in February 2025.
W5
Why
It targets the voice presence problem, aiming to make synthetic conversational speech sound less flat than typical TTS.
W6
With
Built on a Llama style backbone plus the Mimi audio codec, runnable via PyTorch.
W7
Watch
Released under the Apache 2.0 license by Sesame; base model open, though the full production voice demo is not fully released.