Project
LiveKit is an open-source platform for real-time audio/video and voice-agent applications, providing an SFU, SIP integration, and client/agent SDKs.
Microsoft's official plugin lets LiveKit Agents stream live audio to MAI-Transcribe-2-Streaming, a realtime model covering roughly 60 languages. The service is still a public preview that Microsoft does not recommend for production.
Reality
LiveKit Agents now speaks in Microsoft's MAI voices, including MAI-Voice-2-Flash, through an official plugin that covers text-to-speech only. Teams no longer write their own adapter, but every deployment still needs an Azure Speech key and region.
Reality
Self-hosted LiveKit keeps SIP trunks in Redis, so recreating a stock Redis container wipes them and reprovisioning returns new trunk IDs. Health checks stay green until an outbound call fails on a trunk ID that no longer exists.
Reality
Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise, which puts a speaking video agent in scope for production. Extended Thinking is still in preview, so anything shipping now runs without it and the team still builds its own CRM and handoff integration.
Reality
Google's two new voice models process speech and reasoning at the same time and can fire tool calls without pausing the conversation. Published scores put the stronger one top of a speech-to-speech index and lowest on banking tasks.
Perspective Coverage
3 publishersReality
One developer's four clean runs against livekit-client 2.22.3 all read the track dimensions after the preview appeared. Reading there lands outside the roughly 100 ms window in which a portrait iPhone claims 1920 by 1080.
Reality
Google's September 15 release pairs a fast speech-to-speech model with one that narrates its own reasoning aloud while tools run. Clients now have to read interactionStatus to know when a turn is actually over.
Reality
When the acoustic path changes mid-playback, the browser's echo canceller stops cancelling and the leftover sound reaches the voice detector as speech. The fix was 300 extra milliseconds, applied only during playback.
Reality
A dev.to breakdown makes the case that turn detection, endpointing and barge-in are the hard parts of a voice product, and a model swap does not touch any of them.
Reality