Nova3 key-term Prompting

[tannypepsales]

We've been using the Key Terms feature for the past month and found it quite effective in detecting domain-specific keywords—it's been a valuable addition to our pipeline. That said, we've noticed an issue where it seems to force-fit the terms provided, sometimes to the point of inaccuracy.

For example:
It converts "Fahad" to "Fathom"
It interprets "Slaes" as "Pepsales"

These substitutions are problematic, especially in contexts where accurate term recognition is critical. We understand the intent behind forcing matches to key terms, but in cases like this, it introduces noise rather than clarity.

If there's a way to reduce this aggressive auto-correction—or ideally, to allow a setting where key terms influence recognition without overriding what's actually said—it would make this feature significantly more usable for us.

Key Term recognition is important for our use case, and we'd love to continue using it—just with a bit more nuance.

Thanks for your great work and looking forward to improvements!

[deepgram-community[bot]]

Thanks for asking your question. Please be sure to reply with as much detail as possible so the community can assist you efficiently.
_{Consider joining our Discord community for more opportunity to engage with your fellow Deepgram users. You can earn points which can be redeemed for cool stuff by being active in our communities!}

[deepgram-community[bot]]

Hey there! It looks like you haven't connected your GitHub account to your Deepgram account. You can do this at https://community.deepgram.com - being verified through this process will allow our team to help you in a much more streamlined fashion.

[deepgram-community[bot]]

It looks like we're missing some important information to help debug your issue. Would you mind providing us with the following details in a reply?

• The programming language you are working in (e.g. JavaScript, Python).
• The deepgram product you are using (e.g Speech to Text, Agent API)
• A request ID that triggered your error or issue.

[deepgram-community[bot]]

Thank you for the feedback! Keyterm Prompting is contextual and model driven, meaning that it does not work in the same way that keyword boosting work where something like intensifiers can adjust relative weighting for these things. If you'd be willing to share examples with Request IDs I'd love to take a look to better inform future iterations.

This message was sent by nick kaimakis from Deepgram, via our community automation.

[tannypepsales]

Here is the Req_id: 27f4f50b-346a-49d7-aab7-dd2bd1084a8f; Account with mail id: ***@***.*** Everywhere it took *"Fahad"* as *"Fathom".* Regards, Tanishk

…

On Fri, 9 May 2025 at 06:29, deepgram-community[bot] < ***@***.***> wrote: Thank you for the feedback! Keyterm Prompting is contextual and model driven, meaning that it does not work in the same way that keyword boosting work where something like intensifiers can adjust relative weighting for these things. If you'd be willing to share examples with Request IDs I'd love to take a look to better inform future iterations. [image: avatar] <https://camo.githubusercontent.com/6494159bb3174699f4e8e2cc61b31cd164e3ae60085bf41ac26f0ac518a2db06/68747470733a2f2f617661746172732e736c61636b2d656467652e636f6d2f323032352d30332d31302f383538303638303036393333335f65386537396531323838336661336663623138385f33322e6a7067> This message was sent by nick kaimakis from Deepgram, via our community automation. — Reply to this email directly, view it on GitHub <#1233 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BQMJ6UFYQQEIEAT3OTZF2PD25P4W3AVCNFSM6AAAAAB4V6NOZOVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMBYGIZDSMY> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[tannypepsales]

Hi nick is there any update on this? Also we're able to find other issue like taking BANT ans Bank(this wasn't part of Key-Term) and other misclassification which are absolutely important for us!

[tannypepsales]

Still waiting for an update?

[tannypepsales]

Hi nick respectfully this is a P0 for us, we cannot wait this much!

[deepgram-community[bot]]

Specific very similar terms like this may be difficult, and you'll need to find the right balance of keyterm prompting. Keyterm prompting is not just a naive 'boost', it's contextual, but you may also want to add these names to your keyterms as well. You can also adjust keyterms across streams by creating multiple streams. Aside from that, we do offer custom model training if you'd like to speak with sales.

This message was sent by nick kaimakis from Deepgram, via our community automation.

[tannypepsales]

We're encountering an issue where increasing the number of keyterms leads to a higher error rate. As more keyterms are introduced, the model begins to overfit, often forcing matches and resulting in increased misclassification. Ironically, the suggested solution of expanding keyterms across multiple streams is contributing to the problem rather than resolving it.

We’d like to better understand the process of custom model training for handling specific keywords, and how this approach might address the misclassification issue we’re seeing with keyterm prompting.

This is a time-sensitive matter — our customers are growing increasingly impatient, so a prompt response would be greatly appreciated.

[deepgram-community[bot]]

I understand, and we really appreciate your feedback and have shared this with our team to dig further into. Because keyterm prompting is model driven, pathologies like this are more difficult to debug and will take some time. Feel free to try out Nova-2 with keyterm boosting instead in the meantime.

This message was sent by nick kaimakis from Deepgram, via our community automation.

[tannypepsales]

We were facing similar issues with Nova-2 and some more as wrong text classification. That's why we went on to Nova-3. Can you provide some ETA on Nova-3?