Hello! I have a call like this one: 3793ada1-0f19-4a10-81fc-73b1f762cc7e where the transcription is incomplete. In this call you see at the end the AI said "let me transfer you..." but the transcription shows "me transfer you..." This is just one example but it happens almost all the time. In general it's either a wrong word or a word missing despite the audio being relatively clear. How can I improve the transcription in general? We use deepgram nova3 as our transcriber and I'm just commenting on english use cases...

To improve transcription accuracy, try enabling keyword boosting in your Deepgram transcriber config. Add a

keywords

array to your assistant's transcriber section to boost recognition of important or commonly missed words:

json
"transcriber": {
  "provider": "deepgram",
  "model": "nova-3",
  "language": "en",
  "keywords": ["let:2", "transfer:2", "you:2"]
}

This increases the likelihood these words are transcribed correctly. Adjust intensifiers as needed and focus on words that are often missed. [More details here.](https://docs.vapi.ai/customization/custom-keywords) Source: - [Custom Keywords Vapi Documentation](https://docs.vapi.ai/customization/custom-keywords)

That list could end up being a bit long. There's no better option than coming up with a list of all possible misses the AI may have?

We suggest using nova-2 as it produces more stable results, but we would like to investigate this issue further. Are the other calls also using nova-3 where it is missing pieces of transcription in the logs?