Sesame CSM-1B

Released March 13, 2025, CSM-1B (Conversational Speech Model) is the open-sourced base of the tech behind Sesame's viral Maya voice demo. Built on a Llama backbone with a smaller audio decoder producing Mimi RVQ audio codes, it generates speech conditioned on both text and preceding audio — so replies carry natural pauses, filler, and emotional tone that respond to the conversation, not just the sentence. Weights sit on Hugging Face (sesame/csm-1b) under Apache 2.0, natively supported in Transformers since 4.52, and the repo has drawn ~14.7k GitHub stars.\n\nThe honest caveats: the base checkpoint isn't voice-tuned (the demo's magic comes from a fine-tuned variant Sesame kept), English-only at release, and reference code wants a CUDA GPU plus gated Llama-3.2-1B access. Usage terms prohibit impersonation and deception — sensible guardrails, but they limit voice-cloning applications. For production voice agents, expect fine-tuning work.