ElevenLabs
iOS Safari voice sessions now play the first agent message reliably, thanks to a proactive AudioContext unlock on first user interaction.1
Voice widget transcripts got cleaner rendering. Emotion and audio tags are styled as pills instead of leaking raw text, when strip_audio_tags is off.2 The T&C modal now respects a top-level disable toggle — per-language preset terms no longer override it when the feature is off at root.
Speech Engine calls return the full response. The JavaScript and Python SDKs (v2.49.1) expose the complete Speech Engine API response, not just the summary, for callers that need the raw payload.3
React Native SDK imports no longer pull in DOM-only APIs. Native builds were silently importing web-only code from the client package; the v1.2.2 fix removes those dependencies so native builds succeed without polyfill workarounds.4
ElevenAgents added text-only conversation filtering and agent version metadata. A text_only filter on conversation list and message search narrows results to text-mode sessions.5 A new endpoint returns version metadata — ID, branch, sequence number, commit timestamp, and access info — for a specific agent version.
The agent platform added auto_resolve_after_inactive_minutes to alerting settings and expanded LLM options with gemini-3.1-flash-lite and qwen35-397b-a17b.
End-of-turn detection and knowledge base attribution remain stable. The agent_response_complete event fires exactly once per turn; used_static_kb_document_ids and used_chunk_ids on transcript turns show which RAG content influenced each response.6 Scribe real-time transcription accepts up to 50 keyterms and a no_verbatim flag for cleaner output.