LiveKit Agents v1.8.12 recovers caller answers GPT-Live never delegated

LiveKit Agents v1.8.12 catches the answers GPT-Live transcribes but never hands to your backend. After a quiet grace period, a detector checks for missed turns and starts the delegation itself.

Slide text

LiveKit Agents Added

LiveKit Agents v1.8.12 recovers caller answers GPT-Live never delegated

GPTLiveModel(delegation_recovery=True)

After a quiet grace period · answers spoken over the agent get a note

GitHub release · livekit-agents@1.8.12 · Delegation recovery

Gemini 4 Carbon and Gemini 4 Hydrogen are now in the Google plugin, and Hydrogen also runs on LiveKit Inference. Google and Perplexity both add decision models through the Decisions API.

Slide text

LiveKit Agents Added

Gemini 4 Carbon and Gemini 4 Hydrogen join the Google plugin

Hydrogen Also available through LiveKit Inference

Google gemini-decision-1 through the Decisions API

Perplexity Decider v1.3, v1.2 and latest, Flash and standard

GitHub release · livekit-agents@1.8.12 · New providers and models

Lokutor arrives with speech-to-text and text-to-speech across nine languages. AssemblyAI adds a streaming model built for real-time agents, and Baseten adds its streaming v2 transcription protocol.

Slide text

LiveKit Agents Added

Lokutor, AssemblyAI and Baseten add speech models

Lokutor Conv 2.0 STT and Versa 2.0 TTS · nine languages

AssemblyAI assemblyai-stt-3-7-realtime-agent streaming STT

Baseten streaming-v2 transcription protocol

GitHub release · livekit-agents@1.8.12 · New providers and models

The local turn detector moves to an updated mini model that also weighs how long the turn has run. xAI speech-to-text on Inference brings word alignment for adaptive interruption.

Slide text

LiveKit Agents Added

Your agent gets better signals for when a caller is done

Turn detector Updated local v1-mini model · reads turn length

xAI xai/stt-2 through Inference · word alignment

GitHub release · livekit-agents@1.8.12 · New providers and models

Two changes can break agents already in production. GPT-Live's delegation-created listeners now also receive backend delegations, and NVIDIA text-to-speech outputs only linear PCM, so any other encoding setting needs updating.

Slide text

LiveKit Agents Breaking

Two changes can break a running GPT-Live or NVIDIA agent

GPT-Live delegation_created now fires for backend delegations too

NVIDIA TTS output is LINEAR_PCM only

Listeners that expect one kind of delegation need a filter

GitHub release · livekit-agents@1.8.12 · Breaking changes

An entrypoint that crashes now reports the job as failed. Manually committed turns survive later voice activity, and the Google plugin needs google-genai 3.2 or later. The release notes on GitHub list every dependency change.

Slide text

LiveKit Agents Changed

Failed jobs now report as failed

Jobs A job closed by shutdown reports JOB_SHUTDOWN

Manual turns Committed turns survive clears and VAD activity

Google Now requires google-genai 3.2 or later

LiveKit Agents v1.8.12 release notes on GitHub

GitHub release · livekit-agents@1.8.12 · Behavior changes