https://vapi.ai/ logo
Join DiscordCommunities
Powered by
AI Interrupting more and more
# support
a
6faa0d66-faff-4ad6-8e86-d19bbab7b072 I'm on holiday (and don't have full access to VAPI); My business partner is complaining the AI is becoming too aggressive and doesn't let people finish what they are saying. Is this a configuration issue or something changed? He swear he didn't touch the configuration.
v
@Alberto It sounds like there might be an issue with the audio-text model's sensitivity, which is responsible for detecting when a user has finished speaking. This could be due to a configuration change or an update that may have affected the model's performance. Since you're on holiday and don't have full access to VAPI, I would recommend checking in with your business partner to ensure that no unintentional changes were made. Additionally, it might be helpful to contact our support team to investigate the issue further and provide a solution. If you'd like, I can guide you through the steps to contact support or provide more information on how to troubleshoot this issue. Please let me know how you would like to proceed!
You can continue this conversation by mentioning me in the message.
a
fab0e8b8-c67b-45e4-9925-d2da25835f17 Another one
s
@Alberto try again now.
a
What have you changed, so I can tell him to do it for every other bot?
if you mean "try again" the ask for help, it still give erroring out 😦
s
Set responseDelaySeconds: 0.4 and llmRequestDelaySeconds: 0.1 and numWordsToInterruptAssistant as 1
a
@Shubham Bajaj thaat would not make the situation worse? Our use case is different from the normal , we do Q&A and we want the people on the other end of the phone to be able to speak for more than 3/4/5 minutes without being interrupted
"AI is becoming too aggressive and doesn't let people finish what they are saying. " Reducing the responseDelay and the llmRequest, would not increase the canche of having the bot even more aggressive?
I've just tested on a demo one, using your setting the bot is even more aggressive, i can't stop talking that is talking on top of me. Imagine that our clients take up to 2-3 seconds to think about the answer in the Q&A round.
m
Joining in on this ticket as, while ours is a different use-case, we are experiencing similar increased interuptions. We are getting reports from quite a few customers and seeing it ourselves. From looking at the logs in Vapi my only assumption is that with the new VAD changes, the LLM request is timed of the "user will stop speaking" event rather than the "user has stopped speaking" event. Because of the old setup we had tried to reduce latency by making our turn time as short as possible. This had made the agressiveness much more apparent. We have found increasing llmRequestDelaySeconds has been the most helpful, but the "user will stop talking" event seems to be a bit more variable. This means that sometimes our responses take far too long. Do you think these issues are linked? If not I will make my own ticket, otherwise is there an option for us to toggle the old VAD setup (or get some confirmation on what trigger starts each timing)
a
Yes i think they are linked, and overall today my experience with VAPI has been the worst. We already have two clients that want to cancel our service. The voice is lagging, sometime is even disappearing out of thing air, then it start to become anxious and speak on top of the customer..... We're a startup and losing a customer it's shit. Vapi please check my account as soon as you can, i know we're not the big fish, but how can we become one if VAPI itself is cutting us down?
s
if that's the case you can increase the values, i suggested these generic values as i was unaware of your usecase.
If that's the case you can use 
assistant.startSpeakingPlan
and 
assistant.stopSpeakingPlan
which help you tune to your use case. also, we have a new model we trained for middle of thought detection, feel free to try in 
assistant.startSpeakingPlan.smartEndpointingEnabled
@Alberto check your DM>
h
Can I ask what you mean by middle of thought?
s
user starts speaking... then gets silent to think about something... after that continues again
h
got you so turning on "smartEndpointingEnabled" is better for that middle of thought detection and therefore not interrupting in those moments?
s
it will be better if you can create a new support ticket an tagme there.
@Alberto can you share the recent call id please with new support ticket and tag me there.
a
sorry for the delay 🙂 I'm on holiday at the moment, so i jump on/off
startSpeakingPlan -> where can i find the documentation for this? our customer uses the phone number, is there a way to setup from the UI? or even from the API from the "assistant"?
b
We do roughly 2000-3000 calls daily and have our LLM request delay set to 0.5 and as high as 0.8 in some cases
a
It's annoying in the chit-chat part, however the rest of the conversation for us are Q, and a long Answer from the customer.
@Daniel
b
In my experience "Response delay" doesn't seem to do anything. LLM request is the main driver in the pause from what I see
a
i agree, changing the Response Delay i didn't notice much difference
Have you tried the smartEndpointingEnabled?
b
We are working on implementing it to test, but have not yet
2 Views
PreviousNext
Powered by