Sesame, the AI company behind the viral voice assistant Maya, has open-sourced its base AI model, CSM-1B. The model, which features 1 billion parameters, is now available under an Apache 2.0 license, allowing commercial use with minimal restrictions.
Designed to generate realistic speech, CSM-1B uses a Llama-based backbone and an audio decoder to produce RVQ audio codes, a method also employed by Google’s SoundStream and Meta’s Encodec. While the model can generate various voices, Sesame notes that it hasn’t been fine-tuned for any specific voice and has limited support for non-English languages.
Despite its capabilities, the model lacks strict safeguards. Sesame simply advises users to avoid unethical applications, such as voice mimicry without consent or spreading misinformation.
However, a test of the model on Hugging Face revealed that voice cloning can be done in under a minute, raising concerns about potential misuse. Consumer Reports recently warned that many AI-powered voice cloning tools lack meaningful fraud prevention measures.
Sesame gained widespread attention in February for its advanced voice assistants, Maya and Miles, which exhibit natural speech patterns, including pauses, disfluencies, and the ability to be interrupted—similar to OpenAI’s Voice Mode.
The company, co-founded by Oculus co-creator Brendan Iribe, has received undisclosed funding from major investors, including Andreessen Horowitz, Spark Capital, and Matrix Partners.
Beyond voice assistants, Sesame is also exploring wearable AI technology, including prototype AI glasses designed for all-day use. These glasses will integrate Sesame’s proprietary models, hinting at the company’s broader vision for AI-driven personal assistants.
