https://vapi.ai/ logo
Join Discord
Powered by
Deepgram Keyword Support Issue
# support
v
I’m currently using the Nova-3 model with Deepgram for speech recognition. Initially, I configured it with English and added specific keywords that need to be recognized. In this setup, keyword recognition works accurately. However, after switching the model from English to multilingual, the same keywords are not being recognized correctly. ❓ Clarification: Does keyword boosting/support behave differently in multilingual mode? Are there any limitations or best practices for using keywords with multilingual models? Any guidance would be appreciated.
a
Hey! Yes, in multilingual mode, keyword recognition can behave differently because the model processes multiple languages simultaneously, which may reduce the accuracy of specific keyword boosting. Best practice is to keep keywords language-specific or continue using a single-language model if precise recognition is critical. I can help you tweak the setup for better results if you want.
@Vijinigiri Gowri Shankar
v
Yeah, that would be great, thanks for your help. My agent needs to support both English and Spanish, so I can’t use a single-language model. Is there any simple way to improve keyword recognition in multilingual mode? @Abayomi Praise heritage @User
a
Absolutely @Vijinigiri Gowri Shankar since your agent needs to support both English and Spanish, the key is to keep keywords language-specific and include common variations or synonyms. Using short context phrases instead of single words can also help the model recognize them more reliably. I can help tweak your setup so it works well in both languages.
v
thanks @Abayomi Praise heritage Currently, I’m passing keywords as a comma-separated list like: 
bessemer, cullman, gadsden, huntsville, medical, services, montgomery, shoals, sheffield, stevenson, tuscaloosa, batesville, conway, fayetteville, jonesboro, obot, texarkana, phoenix, tempe, tucson, centennial, colorado, denver
.... How should I change this to follow your suggestion? Should I group keywords by language (English/Spanish)? Do I need to add variations or synonyms for each term? How do I use short context phrases instead of single words in this case? Any example format would really help.
could you please loo into that
s
Hi, apologies for the delay.
there are two compounding limitations at play here. Why keyword boosting breaks in multilingual mode The 
keywords
parameter only works with Nova-2, not Nova-3. For Nova-3 you need to use 
keyterm
instead. However, even 
keyterm
(Keyterm Prompting) doesn't work when multilingual mode is enabled. This is a Deepgram limitation — neither boosting method is supported in multilingual mode, which is why switching from English to multilingual broke your keyword recognition. Two options depending on your needs: Option A — Switch back to English-only (recommended for your use case) Since your keywords are all English location names, use 
language: "en"
with Nova-3 and pass them as `keyterm`:
Copy code
json
{
  "provider": "deepgram",
  "model": "nova-3",
  "language": "en",
  "keyterms": ["bessemer", "cullman", "gadsden", "huntsville", "medical services", "colorado springs"]
}
Note: 
keyterm
supports short phrases (not just single words), so you can group related terms like 
"medical services"
together.
Option B — Keep multilingual, handle at the LLM level If multilingual is required, add the location names to your system prompt and instruct the LLM to fuzzy-match them even if transcription is slightly off:
Copy code
The caller may mention these locations: bessemer, cullman, gadsden, huntsville, montgomery... 
Match the closest location even if the transcription is slightly imprecise.
For your use case, Option A is the cleaner solution since your keywords are English.
PreviousNext
Powered by