Radar ·
Radar: Ava demo brings real-time voice to the browser
Published on November 1, 2026, the demo presents a voice assistant that listens and responds in the browser, with on-device processing and WebGPU execution.
On November 1, 2026, a demonstration of Ava was published, showing an assistant that listens and responds by voice in real time in a browser. According to its description, inference takes place on the user’s device. The stack combines the LFM3-Omni model for voice understanding and generation, WebGPU kernels, and TalkingHead with HeadAudio to synchronize the avatar’s lip movements.
The material presents an open-source speech-to-speech stack running in the browser with WebGPU. The on-device processing claim applies to this demonstration, not to other systems. To check the result, consult the original publication and verify its stated scope, components, and demonstration conditions. If using AI to study or apply the technique, do not submit confidential organizational data without first checking the policy approved for that use.
Original source: Ava: assistente de voz em tempo real no navegador, executado no dispositivo