Sesame, the startup behind the viral virtual assistant Maya, releases its base AI model

RVQ refers to “residual vector quantization,” a technique for encoding audio into discrete tokens called codes. RVQ is used in a number of recent AI audio technologies, including Google’s SoundStream and Meta’s Encodec.

CSM-1B uses a model from Meta’s Llama family as its backbone paired with an audio “decoder” component. A fine-tuned variant of CSM powers Maya, Sesame says.

“The model open-sourced here is a base generation model,” Sesame writes in CSM-1B’s Hugging Face and GitHub repositories. “It is capable of producing a variety of voices, but it has not been fine-tuned on any specific voice […] The model has some capacity for non-English languages due to data contamination in the training data, but it likely won’t do well.”

It’s unclear what data Sesame used to train CSM-1B. The company didn’t say.

It’s worth noting the model has no real safeguards to speak of. Sesame has an honor system and merely urges developers and users not to use the model to mimic a person’s voice without their consent, create misleading content like fake news, or engage in “harmful” or “malicious” activities.

I tried the demo on Hugging Face, and cloning my voice took less than a minute. From there, it was easy to generate speech to my heart’s desire, including on controversial topics like the election and Russian propaganda.

Consumer Reports recently warned that many popular AI-powered voice cloning tools on the market don’t have “meaningful” safeguards to prevent fraud or abuse.

Sesame, co-founded by Oculus co-creator Brendan Iribe, went viral in late February for its assistant tech, which comes close to clearing uncanny valley territory. Maya and Sesame’s other assistant, Miles, take breaths and speak with disfluencies, and can be interrupted while speaking, much like OpenAI’s Voice Mode.