Sesame, the startup behind the viral virtual assistant Maya, publishes its base -ai model

The AI ​​company Sesame has published the basic model, the Maya, the impressively realistic voice assistant.

The model, which is 1 billion parameter of size (“parameter” that relate to individual components of the model), is under an Apache 2.0 license, ie it can be used commercially with a few restrictions. The model is referred to as CSM-1b and generates “RVQ audiocodes” from text and audio entrances, as the description of sesame on the AI ​​DEV Platform compartment on the face of Sesame.

RVQ refers to “residual vector quantization”, a technique for coding audio in discrete tokens, which are referred to as codes. RVQ is used in a series of new AI audio technologies, including Soundstream and Meta -Codec.

CSM-1B uses a model from the Meta Lama family as a backbone, paired with an audio decoder component. A finely coordinated variant of CSM makes Maya, says Sesam.

“The model open here is a model of basic generation,” writes Sesame in CSM-1BS hugging facial and github repositors. “It is able to produce a variety of voices, but it was not coordinated with a certain voice […] The model has due to the data contamination in the training data, but it will probably not do well. “

It is unclear which data sesame CSM-1b was trained. The company didn't say it.

It is worth noting that the model has no real protective measures to speak. Sesam has an honorary system and only asks developers and users not to use the model to imitate the voice of a person without their consent, to create misleading content such as fake messages or to participate with “harmful” or “malicious” activities.

I tried the demo to hug my face, and it took less than a minute. From there it was easy to create speech in the desire of my heart, including controversial topics such as the choice and the Russian propaganda.