Vapi send partial transcription to custom LLM
# supportm
Marco Falcone
02/04/2025, 3:39 PMHi, I open a new ticket on this topic.
In my assistant ad661bd8-d418-4408-ba01-c2d4825740d4 I receive multiple time the same input with incremental transcription.
For example in call 259bcdcb-9203-4a44-bb89-13c388a2452d my server receive the attached JSON bodies.
As you can see both inputs have the same number of messages, but the last message of the first request has as content a substring of the last message of the second request.
Is there a way to remove this behaviour or to understand if the first input is just a partial message?
Thanks for your support
https://cdn.discordapp.com/attachments/1336360452208332831/1336360452418175060/first.json?ex=67a38628&is=67a234a8&hm=c6795dd4963812c7b2c3cd010f2b8469bef5d699d7921af995a1ea12c0d97b56&
https://cdn.discordapp.com/attachments/1336360452208332831/1336360452938010654/second.json?ex=67a38628&is=67a234a8&hm=43a78db1f2bafccacaa7133206bba42848864e1aa52b1e947ae9831727347d70&
v
Vapi
02/04/2025, 3:42 PMmessage has been deleted
Vapi
02/04/2025, 3:42 PM@Marco Falcone
Hello! It sounds like you're experiencing an issue with incremental transcriptions where partial messages are being sent multiple times, with each subsequent message containing more of the complete sentence.
To address this, you can utilize the
assistant.voice.inputPunctuationBoundariesinputPunctuationBoundariesinputPunctuationBoundariesinputPunctuationBoundariesVapi
02/04/2025, 3:42 PMYou can continue this conversation by mentioning me in the message.
m
Marco Falcone
02/04/2025, 4:10 PM@Vapi there is no punctuation in received input
v
Vapi
02/04/2025, 4:10 PM@Marco Falcone
I understand that the lack of punctuation in the received input makes it challenging to determine the completeness of the message. In this case, you might consider implementing a timeout mechanism on your server to wait for a short period after receiving a message to see if a more complete message arrives.
Alternatively, if the incremental transcription is not necessary for your use case, you could check if there's a setting in your voice assistant's configuration to disable incremental transcriptions altogether.
If these options are not suitable or if you need more specific guidance on how to implement them, please let me know, and I'll be happy to assist you further.
m
Marco Falcone
02/05/2025, 9:31 AMHo can I disable incremental transcription with deepgram on Vapi?
s
Shubham Bajaj
02/05/2025, 8:27 PM@Marco Falcone Looking into it.
m
Marco Falcone
02/10/2025, 9:17 AMHi @Shubham Bajaj , any news?
c
Chiranjeet Mishra
02/10/2025, 9:18 AMHey Marco, it is to be expected as we send the request when we think that user has stopped speaking. What you can do it increase the waitSecond to 0.8 it will significantly reduce this issue.
m
Marco Falcone
02/10/2025, 9:36 AMHi @Sahil I already tried this solution but it doesn't solve the issue. Is there a way to recognize if the input message is partial or complete? This should solve the problem
c
Chiranjeet Mishra
02/10/2025, 9:37 AMUnfortunately, for now there is no way to recognize it.
s
Shubham Bajaj
02/10/2025, 10:58 AMHey @Marco Falcone we have checked this carefully and it's a pure endpointing problem. So you either need to increase the wait time-out seconds or using custom endpointing. It is not related to partial transcripts.
Shubham Bajaj
02/10/2025, 10:59 AMAs @Sahil has mentioned you can use 0.8 a good median number for waitSeconds or use custom endpointing which works on phrases based rules.
m
Marco Falcone
02/10/2025, 11:00 AMCan you provide me a link to the documentation for the custom endpointing?
s
Shubham Bajaj
02/10/2025, 11:06 AM@Marco Falcone For now it's not in docs but if you find it usefull will push it to the documentation https://gist.github.com/bajajcodes/cd664c81344ef9f27a4d304de43bd313
3 Views