Keyterm prompting costs $0.05/hour at two vendors: 60 hours is $3.00

ElevenLabs lists Scribe keyterm prompting at $0.050 per hour and AssemblyAI lists keyterms prompting for asynchronous audio at +$0.05 per hour. For 60 hours of product calls, either add-on costs 60 x $0.05 = $3.00 on top of the base transcription. Sume STT takes a language hint but no keyterm list, so there is no equivalent line, and it is priced at $0.01 a minute, or $36.00 for 60 hours.

Rates were read on each vendor's page and on Sume's catalog on 2026-10-09.

What keyterms are for

A keyterm list tells the model which uncommon words to expect, such as product names, people's names or jargon, so it spells them correctly instead of guessing a common word that sounds alike. For a support team that talks about a product with an invented name, the gain is fewer corrections afterward.

What to do with Sume STT

Sume's STT request has audio_url, language_code, duration_seconds, a segmentation block and the usual job fields. Provider knobs are fixed server-side, so you cannot send a keyterm list. If the proper nouns matter, correct them after transcription with a find-and-replace table built from the common misspellings you see in a sample.

When the transcript feeds captions and you already have the intended script, the video caption job offers script_text. Sume keeps the speech-to-text word timings and aligns the burned-in text to your script, so product names appear as you wrote them. That route needs the script, which a support call does not have.

Is $3.00 worth it

On 60 hours, $3.00 is trivial next to the cost of reading the transcripts. Test on one hour first: transcribe it with and without keyterms and count the corrections. If the add-on saves more than a couple of minutes of editing per hour, it has paid for itself at any normal wage.

The more useful comparison is base rate. On these pages Sume's price per hour is the highest of the four rows, which is why this post is about what you give up, not about a saving.

• Vendors with keyterms: $0.05 per hour on top of the base price.
• Sume: no keyterm field; fix names after the fact or align to a script.
• Always test with one real hour before committing 60.

Building a keyterm list

A keyterm list works best when it is short and specific. Collect the 20 or 30 words your team gets wrong most often: product names, feature names, competitor names, and any customer names you are allowed to store. Do not add common words, since that adds noise without helping.

Review the list monthly. A new product launch changes the vocabulary of every call, and a stale list can bias the model toward a name that no longer comes up. Keep the list in version control next to your transcription script so a change is visible to everyone who relies on the output.