Skip across the chapter
Quota Playbook

Transcript word boosting: 4 checks before calls

Quota Playbook Editorial Team · Published · 10 min read

Before boosting words in call transcripts, check the combined list size, phrase separators, malformed values and streaming updates.

Key takeaways

  • measure the whole list before adding terms, using Deepgram's documentation.
  • check the separators in each multi-word phrase before saving it.
  • test each new value on a short call rather than relying on the absence of an error.
  • after a failed update, compare the intended list with the keyterms actually in use.
  • use repeated errors from actual calls to choose entries, following Gong's guidance.

Why wrong words in transcripts derail a call review

A call review depends on the transcript being readable. When a product name, a competitor name, or an acronym comes out wrong, the reviewer has to guess what the rep said, and that guess can shape the coaching point.

Sorting wrong words into groups helps a reviewer tell a repeated pattern from a one-off slip. Gong lists product names, competitor names, acronyms, and industry terms as the kinds of terms to add when they are being transcribed incorrectly, according to Gong. A pattern that repeats across calls is the kind that justifies a list entry.

Before you add terms: pick words that are consistently wrong

Gong's documentation says to use observed mis-transcriptions from actual calls rather than guessing common variations, according to Gong. A single odd transcription does not establish a pattern.

A phrase list is a list of words or phrases you provide ahead of time to help improve their recognition, according to Microsoft. The same source says that when you add a phrase to a phrase list, you increase its importance, so the system is more likely to recognize it, according to Microsoft. That describes Microsoft's phrase list, so confirm your own tool's documentation before assuming the same behavior.

Spell each entry the way your team writes it in its own materials, since a near match is not the same spelling. Check that spelling against a current product page before you save the entry. A product name that the transcript splits into separate words needs its own entry, so the list records what the transcript shows.

Before you change any setting, read the documentation for the tool your team uses and compare each limit against that page rather than a colleague's summary. Keep a note of the page and the date you read it, because the note keeps the check repeatable when the tool changes. Use these pages for the relevant checks.

Publisher Source Review action
Deepgram Keyterm Prompting docs Compare the saved list with the intended list.
Gong Manage custom vocabulary with AI Transcriber Match proposed entries to observed errors.
Microsoft Improve recognition accuracy with phrase list Check which phrase-list behavior applies to your tool.

Check one: the combined list cap

The cap counts the whole list

The first limit is a ceiling on text units. The maximum number of these units across all keyterms is 500, according to Deepgram. Because the wording says all keyterms together, you read the figure as a ceiling on the whole list, not on each entry.

The limit measures text units, according to Deepgram, so a count of entries does not measure it directly. A list of short product names and a list of long industry phrases can have the same number of entries and still use different shares of the figure. Measure the list in those units, and treat any entry count as a rough guide.

Before you add any term, copy the working list into a plain document and mark the long phrases apart from the single words. Multi-word entries may use more of the figure, so review them first. Then check the marked total before adding the next term, not after.

When the list nears the figure, rank the entries by how often each error repeats across calls. Keep the top entries and hold back the rest until the next review, so a new term has to earn its place against the entries already on the list. Show the rep the calls where the error appears in the transcript.

Finish by reading the final list against the figure once more, because a late addition can push the total past it. Write the date beside the final list so a later reviewer can tell which version was checked.

Check two: multi-word phrases

Multi-word phrases need their own step because the separators matter as much as the words. Test each phrase on its own before combining it with the rest of the list. When a transcript looks wrong, you can then see which phrase is responsible.

Join each multi-word phrase with %20 or +

To boost one multi-word phrase, join the words with %20 or +, according to Deepgram. Write each phrase in that joined form before it enters the list, and do not mix joined and spaced versions of the same phrase. Keep a copy of the spaced phrase beside it so the transcript still reads naturally to you.

A short product name and a longer industry phrase should follow the same joined form, so the list follows one rule throughout. If a transcript still misses a phrase, check first whether it was entered in the joined form before you edit other entries. That check looks at one entry, so the rest of the list stays stable while you test.

Check three: some values fail silently

The API accepts a value such as keyterm=term:0.15 and treats it as a single literal keyterm, so it silently boosts nothing rather than failing, according to Deepgram. A missing error message therefore does not prove the boost is working. Treat the absence of an error as a reason to test, not as confirmation.

A quick test shows whether a value takes effect. You add one phrase, run a short call where the phrase is said clearly, and read the transcript for it. If the phrase does not appear as expected, you check the value before adding more terms.

Check four: updates during a live stream

Keyterms can also change while a stream is running, which adds a second point where the limit applies. A list that passed the check before the call may need a fresh look when a change is sent mid-call. Read the documented behavior before you send any change during a call.

Over-limit updates keep the previous keyterms

An over-limit update sent in a Configure message during a "Nova-3" stream returns a ConfigureFailure with code CONFIGURE_KEYTERM_LIMIT_EXCEEDED, and the stream keeps its previous keyterms, according to Deepgram. Read that as a failed change, with the earlier keyterms still in effect.

A team using a different model should check that model's documentation before assuming the same behavior. Keep the source for each model in your review log, so a later reader knows which rules applied.

Before you send a live change, check the revised list against the documented ceiling and count the combined total, according to Deepgram. Settle the revised list in a test before the call starts, not while the rep is talking.

After a failed update, compare the saved list with the list you meant to send before deciding which terms need another test.

After the change: confirm the next calls read better

Once the list is changed, check the next calls rather than assuming the change worked. A single improved transcript is a weak signal, so look for the same product name across several calls before calling the change a success. Several calls give a fairer picture, because one call can be unusual for reasons unrelated to transcription.

Compare the transcript of each call containing the same product name before and after the change, reading the exact spelling each time. Note the spelling in a column beside the earlier version, so the difference is visible at a glance.

If the same term still reads wrong, check the joined-phrase form and the combined total first, before any new word joins the list. A term that stays wrong after both checks may need a different kind of fix, so record that case for the next review.

Keep a log of each term you added, with the date and the call where you first saw the error. Each row can hold the spelling in the transcript, the spelling the rep used, the source of the limit that applied, and whether the term was kept. Use timestamped call notes to connect the error to its place in the call.

Use the log at each review so the list reflects what your team actually hears on calls. Remove entries that no longer appear in calls, which keeps the list short enough to check against the combined ceiling.

Illustrative example: test a repeated error

Review 10 calls and find the same phrase transcribed wrongly in 5. Check its spelling, prepare the joined form, and measure the revised list before saving it. Then test the phrase on a short call and compare the transcript with what was said. These illustrative counts help organize the review; they are not an accuracy threshold.

Use sales call review notes to prepare the coaching discussion after checking the transcript.

Your next step is to pull three recent calls, list the words that are consistently transcribed wrongly, and check that list against the combined ceiling before you add any of them.

Questions

What is the keyterm limit?

Check the combined list size rather than the number of entries. Use the ceiling in the list-size section above and measure the final saved version, according to Deepgram. Keep that version with your review notes so another reviewer can repeat the check.

How do I boost a phrase of several words?

Prepare the phrase in the joined form described above before testing it. Compare the saved value with the intended spelling, using Deepgram's phrase guidance. Keep the readable phrase beside the saved value so you can compare both with the transcript after the call.

What happens if a keyterm value is malformed?

Check the saved value and test the transcript before adding more terms. Use the silent-failure example above to identify the documented problem, according to Deepgram. Keep a copy of the value you tested so a later reviewer can trace an unchanged transcript to that exact entry.

Which models support Keyterm Prompting?

Keyterm Prompting is available for both monolingual and multilingual transcription using the Nova-3 Models, as well as Flux, according to Deepgram. For another model such as Nova-2, use the Keywords feature, according to Deepgram. You check which model your tool runs before choosing between the two.

What happens to keyterms when an update goes over the limit?

Treat the rejected update as a failed change and inspect the keyterms in use. Compare the response with the streaming-update behavior above, according to Deepgram. Keep the attempted list beside the saved list so you can see what needs changing before another test.

Which terms should I add?

Choose entries from repeated errors you have observed in your own calls. Use the categories discussed above to organize the review, following Gong's guidance. Keep the call showing each error beside the proposed spelling so you can explain why the entry belongs on the list.

Sources

  1. Multicurrency deals: 5 value checks before coaching
  2. Get back an old coaching doc version: 4 tools to check
  3. CASL email opt-outs: 3 checks for sales managers
  4. Airtable percent fields: 3 checks before a win rate report

More plays