Drop off

[near siren]

Can you pleas share a conversation id where it dropped off? You can find it in the history.

[regal hull]

I'm looking in the history, and here is the strange thing. It doesn't show anywhere in the transcriptions where I am asking it if it is there. However, in the app, when it happens. I can see that it can hear the micrphone because that shows in the graphic interface. I am going to replace it again with something very straightforward and send you that conversation history. That way I can verify that it went dead in that particular test but did not show up in the transcription.

[near siren]

Could be the last user message before the drop off occurred would be missing. Thanks!

[regal hull]

It always shows the last message as coming from the system but none of my responses after that. I'll replicate, but I've checked numerous conversations. Give me a bit of time to run a few diagnostics and to replicate.

[regal hull]

I was planning on testing it with various languages and prompts, but it happened on the first try. I made it 10:44 before it dropped dead. This is the same as before. I respond and talk, and I can see the app is listening to my microphone, but nothing happens. 0fg7fAsTVzpmAZdkkGEs

nothing that I say after the chat drops is recorded in the transcription. It just goes entirely dead

[near siren]

Thanks will investigate? Did it ever occur at like 1-2 mins or always took like 10mins or more?

[regal hull]

yes. It has happened between 2 and 10 minutes. I haven't been able to have conversations much longer than that. Typically, it happens after 3-5 minutes.

[near siren]

Any LLM?

[regal hull]

It seemed worse with 4o. I was able to have longer conversations with 4o mini. I haven't had a chance to test the remainder of the LLMs. But I've tried multiple headset, multiple microphones, and this machine is wired directly into a gigabit fiber line

[near siren]

Oki! Will check

[regal hull]

If there are any other settings you all want me to test, just ask. I can keep testing

[regal hull]

I gave it a couple of days and decided to test again. Same thing on the first attempt. I made it 3 minutes and 9 seconds, and then it just went dead. It feels like maybe it's some sort of timeout, as the system is saying what the GPT has written, and somewhere in that it takes too much time between reading and response (if one thinks before talking). F4xIMQcKXPP1mv3Pprtj

This is the same scenario where on the screen it shows "Listening" and I can tell that it can hear the microphone because the waveform shows in the graphics. But nothing happens. I keep talking and waiting, but total silence

If anyone has a clue or an update, I can spend a few hours this weekend testing it. The longer the conversation, the better for the prompts that I'm creating

I changed the model to 4 turbo. Same result. This time I made it 3:36 2qHAEjwvcoDHWB0vExeX

In this case, I had it quizzing me when it went dead. Quick question and fast response, but it doesn't respond to anything that I say after that point

[regal hull]

Is there anything else that you'd like me to do or test with?

[near siren]

I checked 2qHAEjwvcoDHWB0vExeX

It seems the connection disconnected on the client hm

Will check if there is any issue on the SDK side

[regal hull]

I eventually hit the disconnect button after it's dead air for a long period of time. I'll test it tomorrow and simply not hit the end button to see what happens.

Are you able to have 30 minute conversations successfully?

[near siren]

Testing now

What browser are you using and version?

Also what max duration of the conversation have you set on the agent?

[regal hull]

i have the max duration set to 3000. I am using Chrome.

chrome version is the latest. All the software is fully up to date

[near siren]

Ok, will try again, OS?

[regal hull]

windows 11, fully up to date with all the latest patches

[regal hull]

I have now tested it on a Windows 11 tablet using Chrome as well as a Samsung Android tablet using Brave. Both have all of the latest updates, and I tested with the systems' default microphone and speakers. Exactly the same random drop problem occurs on both devices. Testing was done using a 3600 second (1 hour) timeout.

and, just to make sure that we are on the same page, the call doesn't drop. The output goes dead and it just sits on listening while you speak into it, with no response.

[regal hull]

I seem to have better luck when changing it to GPT-4 turbo. Maybe there is something causing a problem with the backend connectivity with GPT-4o and 4o mini?

[regal hull]

I set one to GPT4 Turbo. That conversation lasted 38:51. I then changed it to GPT4o mini, and that lasted 1:56.

[near siren]

Interesting!

do you have the id for the gpt4o mini one?

[regal hull]

EWRHeK2FEW8SarnNeg2F <-- and in this scenario, I started the conversation and it began, but then it went dead on the receiving end. I gave it one minute for a response of any type, but it was dead air. Then I asked it, "Can you hear me?" which shows up near the end of the call. Do you think perhaps there is a symbol or type of output coming from the AI that is tripping up the audio output such that it halts the audio output entirely? (random bullet points, random numbering, using copy code format, etc.). Actually, I'm looking at the transcript, and the transcript does not show my voice at the 1:45 mark asking "Can you hear me?" There were two things that I said, once at the 47 second mark and the other at the 1:45 mark, and neither shows up on the transcript, but both can be heard clearly in the audio

[regal hull]

I decided to test it again. The same thing happened. This time, the system responded audibly near the end, but that didn't even make it into the transcription. XpB84QyzVwvK8h3NbKc1

I also doubled my microphone input volume, just to see if that made a difference. It did not

[regal hull]

I'm working a lot in text with 4o right now (not via the conversational AI), and it is returning a lot of random things in markdown language, sometimes with copycode boxes, even for things such as quotations. Will this type of output trip up the conversational AI programming such that it halts the interaction???

[regal hull]

Every conversation that I have still drops off eventually. It seems to happen faster, if I talk over the output and interrupt it.

[near siren]

Thanks for the feedback, we will look into the stability issue before EOW!

[regal hull]

awesome. I'll run some extended tests next week. I'm hoping that the stability can at least achieve a 1-hour conversation, at which point some of the longer and complex prompts can really come into play.

[pastel flicker]

Just to bump this, I am noticing very similar behavior. It seems to be after I have a chat where the user speaks for more than 30 seconds in a row. After than the websocket stops sending back any events expect for pings. It seems to be happening exactly when the user speaks more than 30 seconds

Is this a known limitation or is there a way for this to be overcome?

[regal hull]

I like the additional insight. I'm waiting to test further once the issue is resolved. So far, all of my complex interactions end with drops, and without any exception

I'm willing to test for hours on end, if there is a potential fix

[near siren]

Great insight, we are suspecting the same internally and should have a fix soon!

[pastel flicker]

That’s great news! Update if you happen to have any updates

[regal hull]

Same. Once there is an update, I'll begin testing

[near siren]

This should be now fixed folks!

[regal hull]

Thanks. I'll start testing it out again

[ocean ferry]

Idk if is the same issue you were having, but im having a similar issue.
Everytime i interrupts AI, it gets muted. how ever i can still log all the conversation. So i can detect that my voice is being heard and the AI is responding, but there is no audio

[near siren]

Can you retry Bernardo?

Also @regal hull now should be even better

[regal hull]

Thx. I'll test today