Sesame Unveils CSM-1B, the Core AI Model Behind Maya Virtual Assistant

Sesame, the innovative startup behind the highly realistic virtual assistant Maya, has made its foundational AI model, CSM-1B, available to the public under an Apache 2.0 license, allowing for broad commercial use.

Sesame, the AI trailblazer that’s been turning heads, just dropped the foundational model behind Maya—their voice assistant that’s scarily good at sounding human. Dubbed CSM-1B, this beast packs a billion parameters and comes with an Apache 2.0 license, meaning you can pretty much take it for a spin commercially without jumping through hoops. How does it work? By whipping up ‘RVQ audio codes’ from whatever text or audio you throw at it, using some fancy residual vector quantization magic to turn sounds into neat little tokens. Yeah, it’s the same tech big shots like Google and Meta are playing with in their SoundStream and Encodec projects.

At its heart, CSM-1B borrows a page from Meta’s Llama series, then throws in an audio decoder for good measure. And here’s the kicker: a tweaked version of this model is what makes Maya tick, Sesame says. But hold up—the open-sourced version is more of a jack-of-all-trades, churning out a variety of voices without being fine-tuned for any one in particular. It’s got some multilingual chops too, though don’t expect perfection; the training data was a bit of a mixed bag, leading to some, let’s say, creative interpretations of non-English languages.

Sesame’s not shy about waving the ethical flag, urging folks to play nice and not use the model to spread lies or stir up trouble. But let’s be real: the safeguards are more ‘suggestions’ than ironclad rules, as anyone who’s tried cloning voices or making the model rant about their favorite conspiracy theory can attest.

The brains behind Sesame? None other than Brendan Iribe, the Oculus co-creator, who’s got the tech world buzzing again. With Maya, they’ve nailed the quirks of human speech—ums, ahs, and all—and even let you butt in mid-sentence. Backed by heavy hitters like Andreessen Horowitz and Spark Capital, Sesame’s not just stopping at voice tech. They’re also tinkering with AI glasses you can wear all day, powered by their own custom models. Because why not?