Sesame CSM

Sesame AI Labs, the company behind the viral Maya voice demo, open-sourced CSM-1B in March 2025 as the base model of its Conversational Speech Model line. Rather than plain text to speech, CSM generates Mimi RVQ audio codes from combined text and audio inputs, using a Llama-style transformer backbone paired with a smaller audio decoder, which lets it produce dialogue speech that responds naturally to the preceding conversation. The weights sit on Hugging Face as sesame/csm-1b under Apache-2.0, and the model is natively integrated into Hugging Face Transformers from version 4.52 onward. Running the reference code requires a CUDA-capable GPU, Python 3.10, and access to the gated Llama-3.2-1B download; Windows users swap in triton-windows for the standard triton dependency. The base release is not tuned to any specific voice and has limited non-English ability, and the usage terms prohibit impersonation and deception. With 14.7k GitHub stars it is one of the most watched open speech releases, used in research on context-aware dialogue synthesis and in self-hosted voice agents.