Amit Desai

Voice AI for agents, devices, robots


I like to create magical voice user interfaces for AI agents, devices, and surfaces.

I believe the best voice user interface combines (1) deep intuition for spoken conversation: psycholinguistics, turn-taking, dialog; and (2) expert application of voice AI: cascaded VAD, ASR, NLP, TTS, LLM, diarization, speech-to-speech, full duplex.

The combo solves VUI challenges for: customer phone calls, smart speakers & displays, TV assistants, wearables, robots, earpods, in-car, hands-free, ambient, and other surfaces.

I'm also interested in AI interaction intelligence: a voice agent's actions should minimize a user cost function. For every turn: should the voice agent act, confirm, or reject?


About Me LinkedIn

I'm keen on all things voice :) I lead Roku's Voice AI program and created its multimodal TV AI assistant. Previously I was Alexa's Principal Voice Designer. I founded Voxify, a voice AI startup that automated support phone calls for United Airlines. I launched a voice mobile app that was Top 5 in iOS App Store. I like language, writing, illustration. MIT B.Sc. Computer Science.

Ideas & work

Selected work


Presented at AI Engineer World's Fair, 2026

An under-invested part of voice AI is behavior under uncertainty: after interpreting a command, should the assistant act, confirm, or stop? I propose OUCH (Outcome User Cost Heuristic) as a decision policy to choose behavior that minimizes expected user cost — balancing wrong actions and redundant confirmation.

Read more

Lamp demo

OUCH talk