Changelog

​

Added

• Voice agent links: createAgentLink, listAgentLinks and deleteAgentLink (@interviewflowai/talking-avatar/server) make, list and delete the links to your private voice agents with your API key, each with an expiry and an optional number of calls. Failed requests throw TalkingAvatarApiError.

​

Changed

• Close code 4002 now also ends a running session when the account runs out of credits or reaches its monthly spending limit (checked every 30 seconds).

​

Added

• Realistic Avatars: photorealistic talking avatars streamed to your page, billed per minute. LiveAvatar (/server) runs the session on your server: TTS audio in any chunking or base64, barge-in with how much was heard, speech events, and one route for the page (handleView), or publishing into your LiveKit room. LiveAvatarView (/live) plays it in the page over WebRTC and reconnects by itself.

​

Changed

• Licensed under MIT (was PolyForm Noncommercial 1.0.0): free for any use, commercial included.
• three is now an optional peer dependency: only TalkingAvatar needs it.

​

Changed

• Retrained viseme model: the “th” in think and three now shows the TH mouth shape (76% of test frames, up from 11%). Overall shape accuracy is 82.3% (was 81.7%).

​

Fixed

• A p/b/m at the start of a word now closes the lips; it was dropped when the model heard it for only one frame.
• The first viseme event could be timed slightly before the audio started.

​

Added

• TalkingAvatar: loads a .glb avatar and lip syncs it to live audio (attachStream) or clips (speak), playing the audio itself.
• An in-browser audio-to-viseme model (~260k parameters, 1 MB), trained on LibriTTS-R.
• JALI-style coarticulation from viseme events to mouth weights.
• Natural blinks and brows that follow speaking, listening and thinking; setStatus("listening" | "thinking").
• The morphTargets option for avatars with other morph names; avatars named h_expressions.*_h are mapped automatically.
• The TalkingAvatarView React component (/react).