New providers and models
- VLM Run: new
livekit-plugins-vlmrunplugin for vision-language chat and System One decisions, includingqwen/qwen3.8-27b. - AssemblyAI STT 3.7 Realtime Agent: supported through LiveKit Inference as
assemblyai/assemblyai-stt-3-7-realtime-agent. - Sarvam:
STTRealtimenow supportssaaras:v4.
Breaking changes
- Boson AI: replace
livekit-plugins-bosonwithlivekit-plugins-boson-ai, importlivekit.plugins.boson_ai, and change the[boson]extra to[boson-ai]. The avatar plugin is unchanged. - Slng: the STT/TTS plugin and
[slng]extra have been removed; migrate to another provider. - Gladia: remove
AudioEnergyFilterimports and theenergy_filterconstructor argument. All audio is now forwarded, preserving transcript timestamps; quiet periods no longer trigger filter-driven recording completion. - GPT Live: inference HTTP refusals now raise
APIStatusErrorrather thanAPIConnectionError. Update exception handlers accordingly.
Behavior changes
- OpenAI decisions: direct
DecisionModel()requests now default togpt-6-luna; the inference name remainsopenai/gpt-6-luna-decisions. - Gemini Live: inference sessions now retain Gemini’s own turn detection to avoid missing replies.
- LiveKit Inference: STT, TTS, and eligible realtime sessions now hand off during gateway drains. Failed WebSocket handshakes expose quota details and respect terminal retry denials. Explicit
providerpinning now warns of removal on March 1, 2027; omit it to allow routing and fallback. - Python: package metadata now allows CPython 3.15; some plugins remain capped below it.
This release also improves channel memory cleanup, avatar receiver lifecycle, Lokutor connection recovery, Anthropic judge requests, and TTS markup handling.
What's Changed
- pin livekit-protocol by @theomonnom in #8543
- fix(proto): set every wire field by hand by @theomonnom in #8544
- refactor: remove gladia energy filtering by @chenghao-mou in #8577
- fix(tts): don't let a self-closing prosody marker swallow the next expr tag by @chuenchen309 in #6449
- Expose TTS fallback controls and activation events (LKINF-495) by @russellmartin-livekit in #7185
- chore: remove slng plugin by @chenghao-mou in #8572
- fix: add model and provider to eou traces and logs by @chenghao-mou in #8537
- fix(inference): keep Gemini Live's own turn detection in inference.RealtimeModel by @shawnfeldman in #8547
- fix(inference): Surface quota details from failed WebSocket handshakes by @shawnfeldman in #8548
- fix: use gpt-6-luna for OpenAI decisions by default by @chenghao-mou in #8580
- plugins/phonic: expose the remaining per-tool options by @arunwpm-work in #8531
- Homepage agent: set memory limit to contain framework bugs by @thomasyuill-livekit in #8447
- feat(inference): add assemblyai/assemblyai-stt-3-7-realtime-agent to STT models by @tinalenguyen in #8584
- fix(lokutor): detect a dead STT connection with WebSocketMonitor by @kushaldave002 in #8554
- fix(anthropic): send temperature/top_p/top_k from extra_kwargs in the request body by @richardphoenix in #8586
- fix(aio): free a closed Chan without the cyclic garbage collector by @zek in #8222
- build: allow CPython 3.15 by @longcw in #8587
- Fix avatar receiver registration ownership and expose cleared segment ends by @Zheng-Changfu in #8289
- fix(aws): cap aws-sdk-transcribe-streaming below 0.13 by @prassu017 in #8591
- feat(sarvam): saaras:v4 on STTRealtime by @neil-sarvam in #8384
- fix(inference): deprecate provider pinning by @shawnfeldman in #8430
- fix(sarvam): keep an explicit base_url when update_options switches model by @neil-sarvam in #8386
- feat(sarvam): request timestamps from the REST STT endpoint by @neil-sarvam in #8385
- feat(inference): hand off STT, TTS and realtime on gateway drain by @shawnfeldman in #8506
- fix(gpt-live): end the caller's turn when the model answers, not on a pause by @u9g in #8137
- feat: add VLM Run LLM plugin by @wbrennan899 in #8341
- rename livekit-plugins-boson to livekit-plugins-boson-ai by @tinalenguyen in #8620
- feat(gradium): surface server warnings in STT and TTS by @hhugo-gradium in #8582
- docs: fix stale test file names in threat model after test consolidation by @detail-app[bot] in #8599
- fix(gradium): STT model delay and clean session end (1/2) by @hhugo-gradium in #7870
- fix(openai): speak one assistant message per Responses response by @longcw in #8588
- fix(voice): start the next agent while GPT-Live finishes closing by @longcw in #8624
- fix(workflows): say email symbols as words when GetEmailTask spells the address by @u9g in #8605
- fix: preserve simulation content unless redaction is requested by @chenghao-mou in #8639
- chore(examples): use AssemblyAI assemblyai-stt-3-7-realtime-agent by @dlange-aai in #8618
- fix(voice): keep extra set in on_user_turn_completed on reused preemptive message by @jp-paper in #8615
- fix(sarvam): keep idle STTRealtime streams open by @neil-sarvam in #8383
- fix: improve detection of stale connection by @davidzhao in #8653
- feat(inference): add gemini-4-neon to GoogleModels by @tinalenguyen in #8656
- feat(google): add gemini-4-neon to ChatModels by @tinalenguyen in #8657
- feat: add manual flush to speechmatics stt by @chenghao-mou in #8655
- livekit-agents@1.8.13 by @github-actions[bot] in #8621
New Contributors
- @thomasyuill-livekit made their first contribution in #8447
- @kushaldave002 made their first contribution in #8554
- @richardphoenix made their first contribution in #8586
- @Zheng-Changfu made their first contribution in #8289
- @wbrennan899 made their first contribution in #8341
- @jp-paper made their first contribution in #8615
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.12...livekit-agents@1.8.13