livekit/agents livekit-agents@1.8.12

on GitHub

Delegation recovery

You can recover caller answers that GPT-Live transcribes but never delegates. After a quiet grace period, a detector checks missed turns and starts backend delegation when needed, with a note for answers spoken over the agent.

GPTLiveModel(delegation_recovery=True)

New providers and models

• Lokutor: Conv 2.0 STT and Versa 2.0 TTS across nine languages.
• xAI: xai/stt-2 through Inference, with word alignment for adaptive interruption.
• Google: Gemini 4 Carbon, Gemini 4 Hydrogen (also through Inference), and gemini-decision-1 through the Decisions API.
• Perplexity: Flash and standard Decider v1.3, v1.2, and latest aliases.
• AssemblyAI: assemblyai-stt-3-7-realtime-agent streaming STT.
• Baseten: streaming-v2 transcription protocol.
• Turn detector: updated local v1-mini model with turn-length input.

Breaking changes

• GPT-Live: delegation_created listeners now also receive backend delegations; filter mode="responses" where needed.
• NVIDIA: TTS output is restricted to LINEAR_PCM; update other encoding configurations.

Behavior changes

• Dependencies: Google now requires google-genai>=3.2, Phonic requires phonic>=0.33.4, and local inference requires >=0.3.0,<0.4.
• Jobs: uncaught entrypoint failures report failure; job-triggered session closes report JOB_SHUTDOWN.
• Manual turns: committed snapshots survive clears and subsequent VAD activity.
• ElevenLabs: dialogue models now honor similarity_boost.

Interruptions, tool delivery, handoffs, telemetry authentication, speech connections, and answering-machine detection are more reliable.