Hi, we upgraded to livekit agents 1.3.6 and notice...
# agentsc
creamy-judge-56458
12/08/2025, 8:46 AMHi, we upgraded to livekit agents 1.3.6 and notice that when using ElevenLabs or LMNT TTS with ‘preemtive generation’ that TTS stream is really stuttering or sounds like it’s really trying to push. Before this we had 1.2.18 and there were no issues with preemtive generation. Is this known by the team by any chance? It happens specifically with Deepgram as STT, but also with Soniox.
q
quaint-ambulance-32873
12/08/2025, 9:01 AMCan i see your AgentSession code so I can set up and see ?
c
creamy-judge-56458
12/08/2025, 9:34 AM@quaint-ambulance-32873
_await_ session.start(_agent_=agent_class,_room_=_ctx_.room,_room_options_=room_io.RoomOptions(_text_input_=room_io.TextInputOptions(_text_input_cb_=handle_text_input,_audio_input_=room_io.AudioInputOptions(_noise_cancellation_=noise_cancellation.BVC(),q
quaint-ambulance-32873
12/08/2025, 9:36 AMDo you have a constructer where you defined all the stt, tts, llm configurations?
c
creamy-judge-56458
12/08/2025, 9:37 AMWe load the TTS, STT and LLM dynamically, so when the session starts the STT, TTS and LLM models are loaded through config that we store, this happens prior to the agent session being started
✅ 1
creamy-judge-56458
12/08/2025, 9:37 AMWhat I notice is that with preemtive generation on, it’s quite random, sometimes it goes ok, sometimes it goes good, but many times it stutters. I also read that this is a known issue with preemtive generation, is that correct?
q
quaint-ambulance-32873
12/08/2025, 9:38 AMLet me have a look into this and get back to you
c
creamy-judge-56458
12/08/2025, 9:38 AMThanks
creamy-judge-56458
12/08/2025, 9:42 AMHmm I am seeing this also happen now without Preemtive generation disabled, need to investigate a bit further, that TTS stream sometimes sounds stuttery
creamy-judge-56458
12/08/2025, 9:52 AMSomehow this issue is new. We have currently agents 1.2.18 deployed on production and staging, no issues there.
But the issue is there with agents version 1.3.6 on our dev environment
✅ 1
q
quaint-ambulance-32873
12/08/2025, 11:45 AMI can not seem to replicate the issues locally (region can also be a factor here, i am based in europe), but I would suggest looking at the observability insight to understance the traces https://docs.livekit.io/agents/observability/,
c
creamy-judge-56458
12/08/2025, 11:53 AMThanks Karison for testing, I am based in EU as well. We have implemented Tempo for traceability ourselves and see nothing there, I have done around 50+ calls just now and somehow it is now resolved. Wondering if it might have been a phone connection issuer
creamy-judge-56458
12/08/2025, 11:53 AMThanks for checking out
q
quaint-ambulance-32873
12/08/2025, 11:57 AMinteresting, if you have any more information, feel free to share them
c
creamy-judge-56458
12/08/2025, 11:58 AMWill do, we are keeping in eye on it
m
melodic-arm-17634
12/09/2025, 1:09 PMWe've had a similar problem, and it's very sporadic. Our users say that it sounds like the AI is speaking in gibberish, or is elongating vowel sounds for a very long time.
I don't have any solutions, and I've reached out to ElevenLabs support, but haven't gotten anything useful so far
Nothing new to add, but just wanted to say that I'm experiencing this as well!