Node.js 1.2.6: No audio from session.say() or generateReply() in onEnter — speech handle created but no frames produced, no errors, agent state transitions initializing → listening (skipping speaking)

[Lugiox]

Describe the bug

Description
When invoking session.say() consistently produces no audio output, despite:
or session.generateReply() • The agent joining the room successfully
• onEnter firing correctly (confirmed via logs)
• A speech handle being created (confirmed via framework debug log)
• Server-side dashboard showing track published by agent for ~21 seconds
• No errors surfaced in logs
• No unhandled promise rejections
• Unchanged across 3 TTS providers and 2 routing paths (Inference + direct plugin)
The agent state transitions directly from initializing from Agent.onEnter(), the agent
to listening, skipping speaking entirely.
Environment
• @livekit/agents: 1.2.6 (deployed as 1.2.6 in LiveKit Cloud)
• @livekit/agents-plugin-inworld: 1.2.6 (also tested with inference.TTS cartesia/openai)
• @livekit/agents-plugin-deepgram: 1.2.6
• @livekit/agents-plugin-livekit: 1.2.4 (turn detector)
• @livekit/agents-plugin-silero: (VAD)
• Node.js 22 (node:22-slim container)
• TypeScript 5.9, Vite 7.3.2 for build, pnpm 10
• Deployed to LiveKit Cloud "Build" plan, us-east region
• Agent ID: CA_gPrJTQK9WYGF
inference.LLM({ model: 'openai/gpt-4.1-mini' }) (placeholder)
Reproduction steps

1. AgentSession configured with:
   • STT: deepgram.STT({ model: 'nova-3', language: 'en' })
   • LLM: • TTS: inworld.TTS({ voice: 'Serena', model: 'inworld-tts-1.5-max' }) (direct plugin)
   • VAD: Silero
   • turnDetection: • preemptiveGeneration: true
   • noiseCancellation: audioEnhancement({ model: 'quailVfL' })
2. Agent class extends voice.Agent livekit.turnDetector.MultilingualModel()
   with onEnter() calling this.session.generateReply({
   instructions: 'greet the user' })
3. Entry function: ctx.connect() → ctx.waitForParticipant() → new
   voice.AgentSession(...) → session.start(...)
4. Dispatch via agents-playground.livekit.io with correct agent name
   Observed behavior
   Agent logs (most recent test with generateReply):
   [Mindi] Connecting to room...
   [Mindi] Connected to room
   [Mindi] Waiting for participant...
   [Mindi] Participant joined: identity=identity-Le19
   [Mindi] Starting session...
   [Mindi] onEnter — triggering generateReply with greeting hint: "..."
   {"level":30,"speech_id":"speech_bd82a69a-6b3","msg":"Creating speech handle"}
   [Mindi] Session started — Agent.onEnter should have fired
   (indefinite silence — no further events)
   initializing → listening (never enters speaking)
   Playground client (Chrome, Safari, DuckDuckGo all tested):
   • Agent STATE CHANGED event: • "Waiting for agent audio track..." persists indefinitely
   • Agent Identity field stays on loading spinner (or briefly populates then reverts)
   Server-side LiveKit Cloud dashboard:
   Event sequence for a representative session (RM_rjYGoRizLkjZ):
   12:08:30.374173 Participant joining: agent-AJ_tQWZZuJbYpjs
   12:08:30.473045 Participant active: agent-AJ_tQWZZuJbYpjs
   12:08:30.689505 Track PUBLISHED: agent publishes audio track
   ...track remains published for 21 seconds with no apparent frames...
   12:08:51.645851 Track UNPUBLISHED: agent unpublishes track
   So the agent DOES publish a track to the room server-side — but apparently with no audio frames —
   before eventually unpublishing.
   What we've ruled out via systematic testing
   Configuration tested in 10 iterations (v1 → v10):
   Variable Tested values Result
   TTS provider inference.TTS(openai/tts-1) Errors loudly: APIError:
   LiveKit TTS returned error:
   undefined at
   inference/tts.js:391
   TTS provider inference.TTS(cartesia/soni
   Silent hang, no error
   c-3)
   TTS provider inference.TTS(inworld/inwor
   Silent hang, no error
   ld-tts-1.5-max)
   TTS provider inworld.TTS (direct plugin, own
   Silent hang, no error
   API key)
   Speech trigger session.say(text) Silent hang, no Creating
   speech handle log
   Speech trigger session.generateReply({
   Silent hang, but DOES log
   instructions })
   Creating speech handle with
   speech_id
   Agent config Custom llmNode override
   (Claude routing with prefills)
   Silent hang
   Agent config Stripped llmNode default)
   (framework
   Silent hang
   Entry ordering session.start() before
   Silent hang
   ctx.connect()
   Entry ordering ctx.connect() →
   Silent hang
   waitForParticipant() →
   session.start() (canonical)
   Browser Chrome (Blink) "Waiting for agent audio track..."
   Browser Safari, DuckDuckGo (WebKit) "Waiting for agent audio track..."
   Dispatch Correct agent name
   Silent hang
   mymindi-spike
   Dispatch Auto-dispatch (blank agent
   name)
   Silent hang
   Not ruled out yet
   • Framework bug specific to this deployment setup (1.2.6 on Cloud us-east)
   • Interaction between noiseCancellation: audioEnhancement({ model: 'quailVfL' }) and
   TTS output (worth testing with this removed)
   • Interaction with custom VAD/turnDetection config during greeting (before user speech)
   • E2EE accidentally enabled at project level (not visible in dashboard but possible)
   Expected behavior
   Agent should enter speaking state, produce audio frames via TTS, publish them to the
   already-published track, audio plays in the client, then moves to agent.onEnter() promise resolves and agent
   listening.
   Impact
   This blocks deployment to production for a voice agent product targeting elderly users. We've spent
   ~14 hours debugging with detailed logs, dashboard traces, and code iteration before filing. Happy to
   provide additional information, deploy test builds, or give a LiveKit engineer direct access to the agent
   for debugging.
   Request
   We'd appreciate help identifying:
5. 3. 1. Why session.say() produces no Creating speech handle log while generateReply() does
         Why the published track contains no audio frames
         Whether there's a known issue with the specific combination of plugins we're using
         Thank you for the framework. Would love to get this working.

Relevant log output

georgechen@Georges-MacBook-Pro mymindi-spike % cd ~/mymindi-spike
lk agent logs --log-type=deploy > /tmp/v10-test.txt 2>&1 &
sleep 12
kill %1 2>/dev/null
tail -60 /tmp/v10-test.txt
[1] 50125
[1] + exit 1 lk agent logs --log-type=deploy > /tmp/v10-test.txt 2>&1
Using project [mymindi]
Using agent [CA_gPrJTQK9WYGF]

agent-starter-node@1.0.0 start /app
node dist/main.js start

◇ injected env (0) from .env.local // tip: ⌘ enable debugging { debug: true }
{"level":40,"time":1776852561500,"pid":47,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","version":"1.2.6","msg":"custom loadThreshold is not supported when deploying to Cloud, using defaults"}
{"level":30,"time":1776852565325,"pid":47,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","version":"1.2.6","msg":"starting worker"}
{"level":30,"time":1776852565370,"pid":47,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","msg":"Server is listening on port 8081"}
{"level":30,"time":1776852565401,"pid":47,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","version":"1.2.6","id":"CAW_S3R96aKnQU4j","server_info":{"edition":"Cloud","version":"1.10.1","protocol":17,"region":"US East B","nodeId":"NC_OASHBURN1B_CzEb25gLovbd","debugInfo":"","agentProtocol":0},"msg":"registered worker"}
◇ injected env (0) from .env.local // tip: ⌘ override existing { override: true }
◇ injected env (0) from .env.local // tip: ⌘ suppress logs { quiet: true }
◇ injected env (0) from .env.local // tip: ⌘ enable debugging { debug: true }
{"level":30,"time":1776852894215,"pid":47,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","version":"1.2.6","jobId":"AJ_QGXBAeb6xChQ","resuming":false,"agentName":"mymindi-spike","msg":"received job request"}
[Mindi] Connecting to room...
[Mindi] Connected to room
[Mindi] Waiting for participant...
[Mindi] Participant joined: identity=identity-Le19
[SessionContext] No metadata found — using TEST_USER fallback
[Mindi] Session context — user=George, firstCall=false, tz=America/Los_Angeles, callId=test-call-1776852566652
[Mindi] Starting session...
{"level":30,"time":1776852894877,"pid":101,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","participantValue":"identity-Le19","trackPublications":[],"lengthOfTrackPublications":0,"msg":"participantValue.trackPublications"}
[Mindi] onEnter — triggering generateReply with greeting hint: "Hey George! What are you up to this morning?"
{"level":30,"time":1776852894937,"pid":101,"hostname":"deployment-p-5nn3sr48n4a-ca-gprjtqk9wygf-5ff48fcd8d-lnvwf","speech_id":"speech_bd82a69a-6b3","msg":"Creating speech handle"}
[Mindi] Session started — Agent.onEnter should have fired
◇ injected env (0) from .env.local // tip: ⌘ custom filepath { path: '/custom/path/.env' }
scanner error: context canceled

Describe your environment

• OS: macOS Tahoe 26.4.1
Project: p_5nn3sr48n4a (mymindi)
• Region: us-east
• Sample session IDs with the silent-hang behavior:
(room playground-s1J2-xNC0)
• Various others — Agent Observability is enabled, so traces are available
• Agent deployment: • Deploy version tonight: v20260422070552 • RM_rjYGoRizLkjZ CA_gPrJTQK9WYGF
(cartesia variant), plus subsequent v9/v10 deploys

Minimal reproducible example

No response

Additional information

No response

[Lugiox]

Update: Root cause identified — noiseCancellation via @livekit/plugins-ai-coustics@0.2.10

After extensive bisection, I've reproduced the silent TTS hang reliably to a single configuration option.

Reproduction

Starting from the LiveKit canonical agent starter template (Deepgram Nova-3 STT + Inworld direct plugin TTS + Silero VAD + MultilingualModel turn detector + basic Agent class with generateReply() in onEnter), audio works perfectly end-to-end: greeting plays, conversation proceeds, graceful goodbye.

Adding ONE thing to session.start() reproduces the silent hang:

inputOptions: {
  noiseCancellation: audioEnhancement({ model: 'quailVfL' }),
}

Where:

import { audioEnhancement } from '@livekit/plugins-ai-coustics';

Failure mode (with noiseCancellation)

• Agent joins room successfully (identity field populates briefly)
• Agent Audio indicator flashes briefly (middle bar lights up red for ~1 second)
• Agent then drops from the room entirely (identity field returns to loading spinner, "Waiting for agent audio track...")
• No errors in the deploy logs
• Log sequence: received job request → [custom] Connecting to room... → [custom] Connected → ... → (silence; no Creating speech handle ever logged)

Success mode (without noiseCancellation)

• Agent joins, greeting speech handle created
• playout completed without interruption in logs ~6 seconds after handle creation
• Full conversation, preemptive generation works correctly, graceful session close

Environment

• @livekit/agents@1.2.6
• @livekit/plugins-ai-coustics@0.2.10
• @livekit/agents-plugin-livekit (latest)
• @livekit/agents-plugin-deepgram, @livekit/agents-plugin-silero
• @livekit/agents-plugin-inworld@1.2.6
• Deployed to LiveKit Cloud, us-east region
• Node 22.18.0, TypeScript 5.9, pnpm 10
• macOS (darwin-arm64), but bug manifests in Linux container (node:22-slim)

Questions

1. Is this a known compatibility issue between @livekit/plugins-ai-coustics@0.2.10 and @livekit/agents@1.2.6?
2. Is there a different model value for audioEnhancement that's recommended?
3. Should noiseCancellation be configured via RoomInputOptions rather than session.start().inputOptions? The docs I found were ambiguous on this.

Happy to provide minimal reproduction repo, full logs, or anything else that would help. Workaround for me in the meantime is to ship without noise cancellation.