We are encountering quite a few problems with the transcripts returned by VAPI which is different from that obtained directly from the provider. This is the configuration we have

{
     'provider': 'deepgram',
     'model': 'nova-2-phonecall',
     'language': 'en',
 }

In this particular case it has transcribed the caller's name by a number

"call_sid": "62210c63-92b9-4a22-9074-cae47b394a39"

[Start: 00:00:15] [Voice Assistant] Let's start with gathering some information for the shift coverage process. Could you please provide your full name? [End: 00:00:21]
[Start: 00:00:24] [Caller] 2 0 0 [End: 00:00:25]
[Start: 00:00:27] [Voice Assistant] Could you please clarify if 2 0 0 is your full name if you meant something else. [End: 00:00:31]

We did an initial test using the full recording and view the transcript obtained by Deepgram playground. And the transcript obtained is consistent with the recording. We thought it might be a streaming problem and with a fragment of the recording it would be different, so we cut out only the part where the name is pronounced and repeated the test. However, the result obtained in DG was still correct and different from the result provided by VAPI. The next test was to do the same test but changing to another model of lower capacity

nova

, and here we got the same result that was obtained during the call. How can this be possible? Can there be a change of model for any reason? https://cdn.discordapp.com/attachments/1338470163825033287/1338470164190068746/image1.png?ex=67ab32fa&is=67a9e17a&hm=05ad1b1430b4348589740e7b72911e61ca8eaabf44112b96c1c62a3d7b023873& https://cdn.discordapp.com/attachments/1338470163825033287/1338470164550783046/image.png?ex=67ab32fa&is=67a9e17a&hm=99126fdcfa57d6aab5184b793a0c3a25156e92dcf26da05475d0ca99f1b7dfe9& https://cdn.discordapp.com/attachments/1338470163825033287/1338470164903235635/image.png?ex=67ab32fa&is=67a9e17a&hm=62f018ae0a912825e04c36dfb4f186d8d1d5f6b71d0d3025ebab1f9cd525758d&

Hey, it is to be expected because when you attach the audio directly it can take a look in future and predict on the basis of it but that's not the case with streaming of audio. Although, setting endpointing to 300 will improve the quality. Also, nova-3 is coming up pretty soon.

But that does not explain how in a live call using the recording to pronounce exactly the same thing I have not obtained in any case the same result, if I set the

nova

model, I have not managed to obtain the value of the transcription above but erroneous results that may be more similar like "Two over, Sean"

Hey, can you please schedule a tech call with my colleague? He can provide you with proper insights about it: