Thanks @omar.venturi @heivoll, we will inform you when we have an update.
Hi @K3 and team,
We’re now approaching one full month with the audio-cut issue still impacting production.
Could you share any updates on your investigation or an ETA for a fix?
The situation is becoming very critical on our side, and I’m ready to help in any way I can:
- run fresh tests in specific scenarios / time windows
- capture new screen recordings
- generate detailed logs or traces
Just let me know what would be most useful.
Thanks for keeping us posted.
We are working on it and plan to release an update tomorrow.
Hello @heivoll @omar.venturi,
Could you please test?
Hi @K3,
I’ve just tested our local build with SDK 3.6.2 and it finally seems to work. 
I did notice one brief audio dropout during a sentence, but the stream recovered right away.
One thing to flag: the lip-sync now feels much more offset than before. In two instances the mouth stopped moving several seconds before the audio finished. (The Extra Params field is empty on my side.)
Tomorrow I’ll run a full live test: I’ll deploy the new build to production and should have solid feedback within 24 hours.
I’ll hold off marking this as resolved until the field test is done, but my first impression is positive.
Thanks for the quick turnaround!
If lipsync data is not received, the mouth won’t move. Previously, the reason for audio cutting out was due to the absence of lipsync data, which prevented audio playback. Now, even if lipsync data is missing, the audio will still play.
Ok, thanks.
Just to understand better—since it’s something my clients will likely ask—what causes the lipsync data to not be received?
Is it something that will be addressed?
Of course, having the mouth stop moving is better than having the audio cut off, but from a user experience perspective it’s still not ideal.
Can you let me know?
Exactly the same issue we are facing, There is audio but no lip sync or lip movement. The full text is not spoken and cuts out. It’s hard to do a screen recording as it is not everytime. Also the lip Sync is not aligned with the text received.
@K3 Thank you for the update! We’re currently testing it, and I will share more results. As of the first few sessions, it seems that audio doesn’t completely stop anymore, but similarly to @omar.venturi and @Jeasy_Sehgal , we also experience them closing the mouth completely at times. And I’ve had a couple of long delays during speech, but that seems to be at the start of a new sentence or half sentence (after a comma).
I have also had two instances so far where I got my game ending abruptly while the character was saying their last line - probably because FinishedTalking was sent early and they had their IsTalking set to false for over a second. This is what we currently use to see if they’re done saying their last line (as there is no event indicating them being completely done delivering their response).
@K3 Some more testing done. All in all, it’s an improvement, but we’re still experiencing quite a few issues. The audio does glitch and stop sometimes, but the end of the convos are not cut; instead, the facial animations are missing. It does help, so we are going live with this change for our customers - but I do hope we’ll see further improvements, it’s still feeling quite unstable.
Glad to hear the quick fix helped!
This was a quick fix, resolving the lipsync issue will take more time. Regarding the facial animation cutoff, we haven’t been able to reproduce the issue on our end.
Thanks again for your feedback and patience!
Hi @K3 and team,
I’ve just received some field feedback from our client after today’s rollout of 3.6.2 (all times below are CET / Italian time):
- Morning (around 09:00–12:00) – The hologram worked fairly well.
- Lips often stopped moving near the end of long answers, but the audio did finish.
- Afternoon (after 14:00) – More frequent issues:
- The character would pause mid-sentence, stay silent for a few seconds, then continue.
- Occasionally it never resumed.
- Resetting the session restored normal behavior for a short while, but after several back-to-back questions the pauses returned.
So, while the quick fix is definitely an improvement, we’re still seeing instability during longer or repeated interactions.
A few reminders / questions
- Custom ElevenLabs voice is essential for our use case; we can’t switch to another TTS provider.
- In May the system ran smoothly for the entire month, so whatever changed happened after that (more or less, after June 10th).
- Could you share any additional details on why lipsync data (or audio packets) sometimes fail to arrive? Knowing the root cause would help us explain the situation to the client.
I remain fully available to:
- run targeted tests at specific times,
- capture new screen recordings,
- generate detailed logs or traces.
Just tell me what would be most helpful.
Thanks again for your effort and let me know how I can assist further!
Note: the character’s ID related to all these notes is 09f85624-eea8-11ef-85b7-42010a7be016
@Jeasy_Sehgal please do not share your API Key. I removed the message.
Ah, didn’t realize that the log I shared had the API key embedded. Thank you for watching out for that @K3.
The lip sync issue is still prevalent. Changed to Llama3-70B with GCP voice and still bad. Better than Chat GPT. Yesterday at an investor presentation meeting, the digital human kept talking but the lips didn’t move. It was very embarrassing. When can this issue be fixed?
Hi @K3 and team,
It’s been about ten days since my last update, and I’d like to share the latest field feedback (all times CET):
Client report (4 days ago)
• Overall stability is better – the avatar now usually finishes its sentences.
• New issue #1 – Mid‑sentence pauses: about once every 6–7 questions (and up to once every 3–4 in the afternoon) the audio stops for a few seconds, then resumes.
• New issue #2 – Lip‑sync drift: the mouth often stops moving several seconds before the audio ends.
So, while we’ve taken a step forward, these two new problems mean the issue is not yet resolved.
Questions for fellow testers
@heivoll, @Juliana, @Jeasy_Sehgal – have you also noticed
- mid‑sentence audio pauses, and/or
- the mouth freezing before the audio finishes
since SDK 3.6.2?
Any workarounds on your side?
Why this is critical
Our installation is in a museum: if the avatar hesitates or its lips stop moving, visitors often abandon the experience after the first interruption.
We have been using Convai successfully from Sept 2024 through May 2025. Since June the quality has dropped, and we urgently need a definitive fix to restore the original experience. A firm timeline for resolution would greatly assist our planning.
Could you please provide an estimated timeline for a definitive fix?
Custom ElevenLabs TTS remains a hard requirement for us.
As always, I’m ready to help: fresh tests, logs, screen recordings—whatever you need.
Thanks for your support.
Hi. I’m on holiday atm, so limited in terms of testing, etc. But I can confirm that we’re also experiencing mid sentence pauses. And also that this issue is critical for us. And, it worked fine for us until early June, which was when the problems started happening.
Have you encountered this problem when using narrative trees? Because sometimes the sections don’t start, or if they do start, the answers are only half there. I didn’t have these problems before.
Nope, I don’t use narrative trees feature.
I have done extensive testing of all the voices with all of the foundational models and found that Claude with GCP voices is the most stable at the moment. Just thought I’d put it here for the benefit of others.