Transcriber is not able to get the spelling of my name properly. Even If I spell my name it gives different spellings every time it tells my name. Example : My name is Kamalesh. The transcriber spells it as Kamlish, Tamalesh, Tamlesh, Kamlesh. Though I ask to correct the spelling through voice it still takes random spelling every time. How to fix this? https://cdn.discordapp.com/attachments/1386701327253508097/1386701327794700439/vapi_transcript.png?ex=685aa9c0&is=68595840&hm=ea8e61fb309e78fe56c315c88e90c0f209c18af4dc01adebf43984f1038e245a&

Which transcriber are you using currently? We may have some suggestions for you

Intially I was using Deepgram. Now changed to Assembly AI. I tried few others as well but I can see the same results

This is most likely a prompt issue. Here I will provide an example prompt:

{
  "assistant": {
    "prompt": "
      You are an expert at collecting names accurately over the phone.
      
      STRATEGY:
      1. Ask: 'Could I get your full name please?'
      2. If unclear: 'Could you spell out your first name for me?'
      3. Confirm each part: 'And your last name is spelled how?'
      4. Read back: 'So that's K-a-t-h-e-r-i-n-e, correct?'
      5. Final check: 'Perfect! I have Katherine Johnson'
      
      PHONETIC ALPHABET:
      Use A as in Alpha, B as in Bravo, C as in Charlie, etc.
      
      FOR COMPLEX NAMES:
      - Ask them to spell it out letter by letter
      - Use phonetic alphabet for confirmation
      - Pay attention to apostrophes, hyphens, or unusual spellings
      - Example: 'Is that O'Connor with an apostrophe between O and C?'
      
      NEVER proceed without clear confirmation of the complete name spelling.
    "
  }
}

Thanks for the reply. Actually it collects the details properly by asking to spell and gets the confirmation. The problem is when the AI tells it again the transcriber gets a different spelling. It is not able to get the exact spelling which was given by the speaker earlier. It takes new spelling every time

Change the temperature of the assistant to a lower value for less variation in assistant speech.

Ok let me try. Thanks for the inputs

Let us know how the testing goes

Sure

If you are still using assembly AI, we suggest you to switch back to Deepgram as it has the highest benchmarks recently for Indian languages. Nova-2 or Nova-3 will work. You can also add common names to the 'keywords' object so if the transcriber detects a name it will already know the correct spelling.

I'm using Deepgram only now. Adding whole set of names to keywords will be difficult. It will be easier if we can add the names or words to the 'keywords' object dynamically during the call

You can dynamically add names to the keywords Nova-2 or keyterm Nova-3 list during each call, improving transcription accuracy without needing a static list..

Thanks. It will be helpful if I can get documentation reference. I'm using web snippet and initiating calls through API. Once I get the name from the receiver during the call I have to update the Keywords during the ongoing call

@kk This functionality doesn't currently exist in Vapi. As a workaround, you can: - Pre-populate keywords with common Indian names in his assistant configuration - Use Deepgram Nova-3 with keyterms for better keyword support

@Shubham Bajaj Thanks for the input. It would be great if VAPI introduce this dynamic updation of keywords as it will improve the transcription accuracy to real good

If you would like to see that feature on our platform, feel free to add it to the feature request list here -->

@Shubham Bajaj I have raised a feature request

Hey kk, I wanted to let you know that we're managing a high volume of support requests at the moment, so our response time might be a bit slower than usual. I truly appreciate your understanding and will get back to you as soon as possible!  Thanks again for your patience!

No problem. Thanks for the assistance