Sesame CSM Freemium

Sesame CSM (Conversational Speech Model) is a voice model that generates speech with the timing and emotional variation of a real conversation, released with open weights on Hugging Face. Sesame’s own Maya and Miles demo voices are free to try in the browser, and the 1B model can be self-hosted.

What is Sesame CSM?

Sesame is the company behind the viral Maya voice demo that convinced a lot of people AI voice had crossed a line. CSM is the model underneath. It pairs a language-model backbone with an audio decoder, so it predicts speech tokens the way a text model predicts words.

The output is context-aware: it handles interruptions, laughs and pauses better than the flat narration you get from older TTS systems. Sesame released the 1B parameter version publicly, which matters because most competitors keep their best voices locked behind an API.

Sesame CSM Features

• Conversational prosody — timing and intonation adapt to dialogue context.
• Voice cloning from a short audio sample.
• Open weights for the 1B model on Hugging Face.
• Real-time streaming so it can talk back in a live loop.
• Multilingual support beyond English in the hosted demos.
• Self-hosting path for teams that cannot send audio to a third party.

Product Highlights

The naturalness is the reason to care. Compare CSM to a plain TTS read of the same sentence and the difference is obvious — CSM breathes, hesitates and shifts pitch. For voice agents, that reduces the uncanny valley that makes users hang up on bots.

Self-hosting the 1B model is not plug-and-play. The weights are public, but you need a GPU with enough VRAM and a working audio pipeline to get decent latency. Sesame’s hosted demo is the smooth path; running it yourself is a weekend project for an ML engineer.

Voice cloning has a consent problem. The model will clone from a short sample, and Sesame’s terms require permission from the person being cloned. Enforcement is on you, and that is a legal risk if you build a product around other people’s voices.

Who Should Use Sesame CSM?

A fit for developers building voice agents, and for researchers who want an open conversational speech model to fine-tune. A poor fit for podcasters who just need a clean narration voice — ElevenLabs and Murf are simpler, cheaper and better tuned for one-way audio.

Sesame CSM Use Cases

• Powering a voice assistant that sounds human on the phone.
• Prototyping an interactive character for a game or toy.
• Dubbing dialogue with emotional delivery rather than flat reads.
• Researching conversational speech models with open weights.
• Building an on-premise voice bot where audio cannot leave the network.

Quick Summary

• Sesame CSM is a conversational speech model from the makers of the Maya voice demo.
• The 1B version is open-weight on Hugging Face.
• Its strength is natural timing and emotional delivery, not narration.
• Self-hosting needs a capable GPU and real setup work.
• Voice cloning requires consent from the person you clone.