Custom vocabulary

Speech recognition is trained on general language. Your product names, your competitors and your internal shorthand are not general language, and a mis-transcribed product name degrades everything downstream: retrieval misses, the suggestion cites the wrong document, the summary reads wrong.

Custom vocabulary is a list of terms handed to the transcription and prompt layers through the call pack.

Adding terms

Settings → Vocabulary. Each entry is a term and an optional boost. Good candidates:

• Product and SKU names, especially invented words or unusual spellings.
• Competitor names, particularly ones that collide with ordinary words.
• Internal shorthand that shows up in discovery ("MEDDIC", "POC", your pricing tiers).
• Customer or account names you say constantly.
• Acronyms your industry uses that a general model would spell out.

Keep the list tight. It is a bias, not a dictionary: a hundred rare terms works better than a thousand common ones, because boosting words that were never going to be misheard only creates opportunities for false positives.

Where terms are applied

Terms travel with the call pack, so they reach:

• Transcription, biasing recognition toward the terms.
• The suggestion prompts, so generated lines spell your product the way you do.

They are resolved together with the team's policy and scorecard rubric in a single call-pack read, which means a team can carry its own terminology without any special-casing in the desktop client.

Checking it worked

The fastest test is a roleplay call: say the term a few times and read the transcript. If it still comes back wrong, try a spelling closer to how it is pronounced rather than how it is written; the bias operates on sound.

API

GET /v1/vocabulary, POST /v1/vocabulary, DELETE /v1/vocabulary/:id.