Duplex Model Support

Speech models that are able to speak and listen simultaneously are now supported via the new DuplexModel class, with OpenAI's GPT-Live model as the first to be implemented.

session = AgentSession(
        llm=GPTLiveModel(
            voice="marin",
            # backend Responses model that handles reasoning and tools
            responses_options={
                "model": "gpt-5.6-luna",
                "instructions": "Use tools when current information is required.",
            },
        ),
    )

Read more about duplex models in our docs.

What's Changed

• chore: isolate ruff dependencies in ci by @chenghao-mou in #7142
• fix(telemetry): describe option objects in the session report by @davidzhao in #7127
• chore: remove commented-out keepalive stub from Deepgram V2 STT by @detail-app[bot] in #7141
• docs: remove stale example references after examples revamp by @detail-app[bot] in #6402
• fix(openai): don't replace a TTS prewarm that is still in flight by @omChauhanDev in #7147
• fix: drain unused adapter metrics streams by @chenghao-mou in #7146
• fix: drop unused audio forwarding buffer by @chenghao-mou in #7144
• fix(tests): run the deepgram handshake redaction check in CI by @omChauhanDev in #7152
• plugins/phonic: add enable_watermarking realtime config option by @arunwpm-work in #7121
• speechmatics: expose end_of_turn_config and vad_config options by @rrfunde in #6945
• fix(worker): discard impossible cgroup v2 cpu samples by @longcw in #7113
• fix(voice): fail remote session requests when transport closes by @EPxiaoma in #7149
• fix(voice): report played audio across an audio sink swap by @longcw in #6967
• fix(agents): don't close in-use connections when the pool is invalidated by @Rehansanjay in #7139
• feat(spatius): warm up region, TLS, and session token at process init by @3DRX in #7025
• fix(asyncai): invalidate the TTS pool when connection params change by @Rehansanjay in #7132
• fix(neuphonic): invalidate the TTS pool when connection params change by @Rehansanjay in #7133
• fix(cartesia): honour api_version on both the pooled socket and the REST header by @Rehansanjay in #7140
• fix: close temporary speech stream adapters by @chenghao-mou in #7154
• fix(hotel_receptionist): don't pass simulation runs the grading never checked by @u9g in #6797
• perf(anam): start engine session with API key by @ao-anam in #6203
• workflows: spell the value back once the caller refuses a confirmation by @u9g in #6990
• fix: name VAD in the slow-inference warning by @claude[bot] in #7176
• feat(meta): add Muse Voice streaming STT plugin by @seyeong-han in #7097
• fix(voice): report e2e_latency for the first reply after a handoff by @longcw in #7167
• fix(tests): run the false-interruption resume tests on virtual time by @omChauhanDev in #7179
• feat(telemetry): flag synchronous code blocking the event loop by @davidzhao in #7128
• fix(llm): preserve Annotated tool argument metadata by @Kuang-xianxin in #7126
• fix(google): skip context replay on resumed realtime sessions by @he-yufeng in #6000
• perf: skip discarded trace payload construction by @chenghao-mou in #7155
• plugins/phonic: make allow_tool_chaining configurable per tool by @arunwpm-work in #7197
• feat(openai): add LiveKit inference realtime model by @russellmartin-livekit in #6035
• feat(plugins-tavus): default to the stock Lucy face when no face or pal is given by @carolin-tavus in #7211
• feat: add DuplexModel for full-duplex speech models by @longcw in #6677
• feat: add openai gpt-live duplex support by @chenghao-mou in #7212
• livekit-agents@1.8.1 by @github-actions[bot] in #7214

New Contributors

• @omChauhanDev made their first contribution in #7147
• @EPxiaoma made their first contribution in #7149
• @ao-anam made their first contribution in #6203
• @claude[bot] made their first contribution in #7176
• @seyeong-han made their first contribution in #7097
• @Kuang-xianxin made their first contribution in #7126

Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.0...livekit-agents@1.8.1