Improve call transcription accuracy

Atender’s voice channel transcribes calls automatically, but a generic speech model doesn’t always know how to spell your product names, industry terms, or your team members’ names. The custom dictionary is a list of words and short phrases the transcription engine biases toward — add the terms it gets wrong, and they start coming out right.

Before you start

• A user role with permission to manage AI settings
• A few examples of words the transcription is currently misspelling — listening to recent call transcripts is the fastest way to find them

Add words

1. Open Settings → Transcription.
2. Type a word or short phrase in the input at the top — for example, Atender, Lucky River, or Erik Olsen.
3. Press Enter or click Add. The entry appears in the list, sorted alphabetically.

Repeat for each term. The dictionary holds up to 1,000 entries, and each entry can be up to 6 words long. Duplicates are rejected.

Remove words

Hover over a row in the list and click the trash icon that appears on the right. The change takes effect on calls that start after the edit.

Find an entry fast

Once you have 10 or more entries, a search input appears above the list. Type any part of the word to filter the list down.

What to add

Good candidates:

• Product names, especially ones with unusual spellings (Atender, Cowork, Anthropic)
• Industry jargon your competitors and customers use that a generic dictionary wouldn’t know
• Names of your team members and frequent customers
• Acronyms callers commonly say in full (MFA, SSO, CSAT)

What’s not worth adding:

• Common English words — they’re already in the base model
• Words that just need capitalization — the dictionary improves recognition, not case
• Long sentences — keep entries to single words or 2–3 word phrases

Verify it worked

Make or wait for a call where the term should come up, then open the call’s transcript afterward. The term should now appear spelled the way you entered it.

Be aware: the dictionary is a bias, not a hard override. If the audio is poor or the speaker pronounces the term unusually, the model can still get it wrong. Repeat occurrences and clearer audio improve the hit rate.

Troubleshooting

• Symptom: “Too many words” error when adding an entry. Fix: Each entry can have at most 6 words. Split longer phrases into multiple entries, or just add the key term.

• Symptom: “Limit reached” error. Fix: You’ve hit the 1,000-entry cap. Remove entries that haven’t helped (especially common-English ones).

• Symptom: A word I added still gets misspelled. Fix: Confirm the spelling in the dictionary matches the spelling you want — case-insensitive, but the exact letters matter. If it’s right and the term still gets missed, the audio quality on that segment is probably the bottleneck, not the dictionary.