wet-aquaW
VAPI2y ago
wet-aqua

Transcriber works well on web but not with twilio

I have created a vapi agent, that is in spanish, it uses deepgram for transcribing into text, azure/11labs for voice, and groq/gpt for the model, right now this works perfectly when being tested on web, as well as the transcripts read perfectly when seen in the "calls" section.

However, when doing the same call, with the exact same settings but on a phone call, the bot acts super weird, it interjects the user mid sentence, randomly stops talking, repeats things, ignores the users message. And this gets reflected when looking at the transcripts.
The transcripts look something like this (note this are long calls with various messages exchanged by the user)

  • USER: Cutted off message of the user
  • AI: Predefined first message
  • USER: Continuation of the cutted off message of the user, followed up by all the other words the user said in the call, all stitched together in the same message
Thats it, no other words are recorded in the logs from the ai, apart from the predefined first message. (even though it does say some things, just cutted off, badly, and off context, and it almost always repeats stuff)

Here are some ids of some calls:
1d32a5e3-2251-4621-af63-e0da4bf98ec7
2de0dd70-d1e9-49bc-87cb-667ed5ef491d
ca54db9d-065a-45ef-8241-fed4798f9d40

Note that i have tried this with 3 different phone numbers, in a quiet setting, and all of them produce the same result

Any idea how this can be fixed?
Was this page helpful?