Sesame CSM

Sesame CSM is an open conversational speech model from Sesame that generates speech from text and audio context. Its GitHub page, read directly, shows an Apache 2.0 licence, about 14,700 stars, 1,500 forks, 9 open issues, 4 open pull requests and 24 commits on main, with the repo public and not archived. The README records a 1B variant released on March 13, 2025 and native support in Hugging Face Transformers from version 4.52.1 on May 20, 2025. The fetched page did not show the date of the latest commit, so I could not confirm recent activity. The company links its voice demo, where a fine-tuned variant powers the interactive demo, and I did not read any Sesame product or pricing page.

The README warns of real limits. The open model is a base generation model that has not been fine-tuned for any specific voice, it cannot generate text, it likely does poorly in non-English, it needs a CUDA GPU and it requires gated access to Llama-3.2-1B and csm-1b on Hugging Face. There is no price because you run it yourself. It is best seen as a research and developer release, not a polished product, and it is useful for building conversational voice prototypes where context improves prosody. Treat any hosted product as separate and check the company site for terms.

Sesame CSM suits researchers and developers experimenting with natural, context-aware speech generation. Compare Kokoro TTS, Fish Audio and Cartesia. Browse the Speech & Transcription tools directory for more.

Conversational speech research

Context-aware prosody.

Voice agent prototypes

Base model.

Is Sesame CSM free?

Yes, Apache 2.0.

Needs a GPU?

Yes, CUDA.

Generates text?

No.

Languages?

English works best.