Sesame CSM

CSM is Sesame AI's Conversational Speech Model, open-sourced in March 2025 under Apache-2.0 after the company's Maya voice demo went viral for its eerily natural pauses and intonation. Unlike plain text-to-speech, CSM generates Mimi audio codes from both text and prior audio, so its delivery responds to the conversation that came before — interruptions, hesitations and all. The 1B-parameter base model sits on Hugging Face as sesame/csm-1b, plugs into Transformers from 4.52, and has picked up around 14.7k GitHub stars.

The honest picture, and it matters: the voice everyone raves about in Sesame's demo is a larger closed production model you cannot download, buy or call — the open CSM-1B is a smaller synthesis block, and people who have run both describe it as distinctly weaker. It's English-only, needs a CUDA GPU and a gated Llama-3.2 download to run, and the license bars impersonation and deception. Treat it as a research-grade building block for voice agents, not a drop-in ElevenLabs replacement.

pip install from the SesameAILabs/csm GitHub repo, or load sesame/csm-1b via Hugging Face Transformers (>=4.52). Needs a CUDA GPU and HF access to the gated Llama-3.2-1B weights.