Sesame, a startup led by Oculus co-founder Brendan Iribe, has unveiled a new AI voice assistant called Maya that aims to cross “the uncanny valley of conversational voice.” According to a recent article by technology journalist Sean Hollister, Maya offers more natural and engaging conversations compared to existing voice assistants like Amazon’s Alexa or Google’s Gemini.
The company claims its Conversational Speech Model (CSM) brings “voice presence” to AI interactions, making them feel more genuine through emotional intelligence, natural conversational dynamics, contextual awareness, and consistent personality traits.
How the technology works
Sesame’s technology approaches speech generation as an end-to-end multimodal learning task. The company has developed a system that:
- Uses transformers to leverage conversation history for more natural speech
- Operates as a single-stage model to improve efficiency and expressivity
- Employs a unique compute amortization scheme to overcome memory bottlenecks
According to technical documentation released by Sesame, their system outperforms existing approaches in objective tests like homograph disambiguation and pronunciation consistency. In subjective evaluations without conversational context, human listeners showed no clear preference between Maya’s generated speech and real human recordings, suggesting the technology has reached a high level of naturalness.
Future plans and company background
Sesame has announced plans to release AI glasses designed to be worn all day, providing “high-quality audio and convenient access to your companion who can observe the world alongside you.” Early prototype images show a sleek design, though few technical details have been shared.
The company is backed by venture capital firms including Andreessen Horowitz, Spark Capital, and Matrix Partners—all previous investors in Oculus VR. In addition to Iribe, the leadership team includes former Ubiquity6 CTO Ankit Kumar and former Meta Reality Labs research engineering director Ryan Brown.
Sesame has committed to open-sourcing its models under an Apache 2.0 license and plans to expand from English to over 20 languages in the coming months. The company acknowledges current limitations, noting that while their system generates high-quality conversational prosody, it can only model the content of a conversation, not its structure.
Those interested can try Maya through a demo on Sesame’s website, which includes a disclaimer that calls are recorded for quality review but are deleted within 30 days and not used for machine learning training.