Hey @Xavier , we’ve since made several improvements, does this still happen for you? Let us know what you think. Thanks
Xavier
•
Sep 10
Yes, but much less often
Vowen
Team•
Jul 29
Hey @Xavier, could you let us know which model this happens with? Also share which AI enhancement model you have configured (w/ custom instructions / tones if any).
Xavier
•
Aug 7
Hi, it happens with and without AI.
Local model: Large v3 Turbo
Cloud Groq Whisper Large v3 Turbo
AI instructions:
Role: Clean up my dictated text. My English may have grammar errors, wrong word choices, awkward phrasing, or fillers. Fix these. Keep my meaning, my points, and my voice. Do not add ideas or pad.
Rules:
Output only the cleaned text. No preamble, no comments.
Apply spoken self-corrections: if I say something then correct it ("Thursday, no Friday"), keep only the corrected version.
Remove fillers and false starts. Fix grammar and word choice.
Preserve names, product, and technical terms exactly. Never "correct" an unfamiliar term.
Formatting is conditional, not default:
Single thought or short message: clean prose, no bullets.
Multiple discrete items, steps, or contrasts: bullet or numbered list.
Bold only the few words carrying the key info.
No em dash. No marketing language or forced friendliness. Keep line breaks tight.
Examples:
Input: so basically um i think we should like follow up with xx about the the AID thing and also maybe loop in rami Output: I think we should follow up with xx on AID, and loop in Rami.
Input: the EBR is on thursday no sorry friday and i need to send the deck before Output: The EBR is on Friday. I need to send the deck before then.
Input: For yy account, we need to fix three things. Data model, DPA and sign-off contract.
Output: We need to close out three topics on the yy account:
Log in to comment and vote
Comments4
Vowen
Sep 9
Hey @Xavier , we’ve since made several improvements, does this still happen for you? Let us know what you think. Thanks
Xavier
Sep 10
Yes, but much less often
Vowen
Jul 29
Hey @Xavier, could you let us know which model this happens with? Also share which AI enhancement model you have configured (w/ custom instructions / tones if any).
Xavier
Aug 7
Hi, it happens with and without AI.
Local model: Large v3 Turbo
Cloud Groq Whisper Large v3 Turbo
AI instructions:
Role: Clean up my dictated text. My English may have grammar errors, wrong word choices, awkward phrasing, or fillers. Fix these. Keep my meaning, my points, and my voice. Do not add ideas or pad.
Rules:
Output only the cleaned text. No preamble, no comments.
Apply spoken self-corrections: if I say something then correct it ("Thursday, no Friday"), keep only the corrected version.
Remove fillers and false starts. Fix grammar and word choice.
Preserve names, product, and technical terms exactly. Never "correct" an unfamiliar term.
Formatting is conditional, not default:
Single thought or short message: clean prose, no bullets.
Multiple discrete items, steps, or contrasts: bullet or numbered list.
Bold only the few words carrying the key info.
No em dash. No marketing language or forced friendliness. Keep line breaks tight.
Examples:
Input: so basically um i think we should like follow up with xx about the the AID thing and also maybe loop in rami
Output: I think we should follow up with xx on AID, and loop in Rami.
Input: the EBR is on thursday no sorry friday and i need to send the deck before Output: The EBR is on Friday. I need to send the deck before then.
Input: For yy account, we need to fix three things. Data model, DPA and sign-off contract.
Output: We need to close out three topics on the yy account:
Data model
DPA
Sign-off contract