https://livekit.io logo
Join Slack
Powered by
Hi guys, We are experiencing that our AI doesn't ...
# telephony
w
Hi guys, We are experiencing that our AI doesn't hear our users. When checking Agent insights (observability), we can't hear the users ourselves, almost like the audio doesn't get through. There is sound from the users although the volume is quite low. We have tested with crisp and without, we have also tested with different phone numbers, airpods, phone models, headsets and all seems to have the same problems. Could it be that our agent doesn't unserstand our users because the track volume is low? This is the room session ID: RM_8Ejkd73dC7nr, for the attached screenshot.
r
Have you checked the PCAP for the call? Who is the SIP provider?
w
Hi @refined-appointment-81829 We use a SIP provider called Puzzel. Also would it be okey to send the PCAP file in DMs, considering it may contain sensetive data, such as phone numbers I would assume
Another thing to note is that this is the observed behaviour when we use our test phone number, which is a livekit bought phonenumber. Would I be correct in my assumption that the volume (in regards to audio loudness) and quality we hear in the agent insight is the same as for the AI agent?
r
You can send the PCAP in DM. Just link this conversation there so I know what it is for.
Yes, what you hear for Agent Insights should be exactly the same thing the agent actually hears.
Is it just this one case that fails and all others work or is this a general issue for your calls
What do you hear if you call in then join that room with a web browser? Do you get same results?
w
It is a recurring issue for most people. about 1/3 are valid
r
I took a look at your pcap and it appear like audio is flowing in both directtions
w
We have an AI bot and a human caller is that what you are referring to?
Also to answer your question about joining a room in web, we get the low volume issue here as well
r
Who is your SIP provider?
w
Our Sip provider is Puzzel.
r
Do they have a gain control in the trunk config?
w
But we also have a livekit bought number, I can find a similar case here if that would help?
r
oh, that is helpful
w
Not sure I would have to ask
awesome, one moment
r
How does the agent sound to the user? Is it quiet or normal
Have you tried disableing noise cancelation?
w
It sounds normal for the user
Yes I have tried turning of BVC and Krisp. Also tested with BVCTelephony, but that did not work either
👍🏼 1
r
w
I'll give it a go now, one moment
r
cool. Also what version of LiveKit are you using? 
pip freeze |grep livekit
w
1.3.5 for the most of it, but I have been experimenting with other versions
👍🏼 1
I also forgot to mention that it feels like it works fine when running the agent in console it is just when It is over telephony. We are also using a realtime model
r
Since you hear the same thing when you personally join the room I think it is independent of agent code now.
w
okey, so it can not be any NC? Could it be my livekit configuration of the SIP trunks or dispatch rule? Should I also contact my SIP provider to ask about the input gain?
r
Well, since you can reproduce with LiveKit number I don't think it is that.
It happens with all callers and not just that one right? When I listen to the recording it sounds like they are speaking very softly to begin with and not just an attenuation issue.
w
It happens to about 60% to 70% of callers, and I agree it sounds like they speak softly, It is me speaking and i talked as I usually do in phone
👍🏼 1
It happens to our end customers as well
👍🏼 1
r
I am checking with our media infra team to see if they have any ideas what might be happening here.
I will let you know once I hear back.
w
Awesome thank you so much!
r
It would be helpful to get the call details for the LiveKit number case. We can look deeper into that one
👍 1
So, you are seeing speaker volume is quiet. I am curious what the impact is. Agent STT should handle audio that is not at optimal volume.
w
Yeah so we have noticed that for the sessions where the audio levels are low the AI can't interpret what the user is saying. It is struggling a bit to begin with, as it is in norwegian, so having that extra layer of volume makes it difficult and frustrating to communicate with the AI agent.
We assume it is due to the volume, as it works fine when the users volume is greater.
👍🏼 1
Open in Slack
PreviousNext
Powered by