← All speakers

Amit Desai

Conference affiliation: Roku

Amit Desai is a voice-interface designer and entrepreneur whose work spans telephone customer service, mobile navigation, and voice AI products at Alexa and Roku. At Voxify, he guided a platform for automated customer conversations; at Dial Directions, he helped build a service that accepted spoken destinations and returned driving directions by text. His work on conversational behavior asks how assistants should respond when they may have misunderstood a user.

On this page

From data tools to spoken interfaces

Amit was co-founder and chief technology officer of Anubis, where his team developed a data-warehouse design tool initially built for consulting clients. Anubis became part of Personify in a 1999 transaction. The 2006 SpeechTEK program connects his work at Anubis and Personify with his subsequent role as founder and chief technology officer of Voxify.

At Voxify, Amit researched voice interfaces and speech recognition and guided the platform behind its hosted call-center self-service. The product used speech technology to handle customer requests over the telephone. His 2006 interview describes responsibilities spanning the underlying recognition technology and the spoken interaction customers experienced.

By 2007, Amit was co-founder and chief product officer of Dial Directions, which made driving directions available through an ordinary mobile phone. A caller spoke an address or store name, then received directions by text message. The service combined speech input with text instructions: callers could request a destination aloud and keep the directions available afterward, without requiring a dedicated GPS device.

Amit later led voice AI products at Alexa and Roku. He was listed as Roku’s Director, Voice & Assistant AI for AI Engineer World’s Fair 2026, extending his work in spoken interfaces from telephone services to consumer devices.

Choosing how an assistant responds

Amit’s Outcome User Cost Heuristic, or OUCH, weighs the effort users spend recovering from an assistant’s behavior. It uses that cost to choose whether to act, ask for confirmation, or stop and request a repetition. Confirmation adds effort even when the assistant understood correctly; acting on a misunderstanding can make recovery harder. His explanation of conversational behavior under uncertainty argues that choosing between these responses can improve an experience without changing recognition accuracy. He extends the concern to assistants that send messages or take physical actions, where mistakes have consequences beyond an inconvenient answer.

Read the topics behind these talks

1 conference talk

Loading more talks and highlights…

References