Sesame AI voice and CSM-1B: the demo, the open model and an alternative
Sesame’s conversational voice demo, with the voices people know as Maya and Miles, impressed many listeners, and Sesame also released an open speech model, CSM-1B. They are not the same thing, and that matters if you want to use the voice in your own project.
This guide explains, according to the CSM-1B model card, what the open model is, what it takes to run, the rules for using it, and when a ready-to-use engine such as SuperClonica is the simpler choice.
The demo and the open model
The model card says a fine-tuned variant of CSM powers Sesame’s interactive voice demo. The released CSM-1B is the base Conversational Speech Model, so it will not sound or behave exactly like the demo out of the box. CSM generates speech from text together with audio context, which is how it keeps a speaker consistent across turns.
What CSM-1B is
CSM uses a Llama backbone and a smaller audio decoder that produces Mimi audio codes. It was released on 13 March 2025 under the Apache-2.0 license and is supported natively in Hugging Face Transformers from version 4.52.1. English is its main language; the card says other languages work only to a limited degree and likely poorly. To download it you must agree to share your Hugging Face contact information.
Running it and the rules
The example code on the card uses a CUDA graphics card when one is available and otherwise falls back to the processor. The card forbids using the model to mimic real people without their explicit consent, to create deceptive content or for illegal or harmful purposes.
A ready alternative on an ordinary computer
If you need speech now rather than a research model, install the free SuperClonica engine with python -m pip install superclonica-cpu and run superclonica --text "Hey, good to hear from you." --voice F1 --lang en --output hello.wav. It runs offline on an ordinary processor, speaks 31 languages, and comes with ten stock voices, M1 to M5 and F1 to F5. The engine is under Apache 2.0 and the Supertonic 3 model under OpenRAIL-M, which allows commercial use within its restrictions.
Voices and your own voice
Hear catalog voices free on their pages: paste 30 to 100 characters, no account needed. For a voice of your own, record a clean phrase of 5 to 10 seconds and train it in the studio; training on your own NVIDIA graphics card is free. Note that the free catalog voice named Maya is SuperClonica’s own and has nothing to do with Sesame.