Capture, transcribe, structure, verify

Dictation involves at least two transformations. Speech recognition converts audio into text. A language model may then organize or rewrite that text. Separating the stages helps you identify whether a mistake arose in transcription or editing. Speech recognition systems have their own model architectures and task choices; transcription quality should be checked on the actual audio conditions. [source]

The purpose is to save typing without losing distinctions you would have preserved in a deliberate written note.

Dictate with light structure

Before starting, state the matter label you intend to use, the document type, and whether the output is an internal note or a draft for someone else. Use verbal markers: “Background,” “Decision,” “Open question,” and “Next action.” Speak names and numbers carefully. For a difficult identifier, spell it or add it manually after transcription.

You do not need to deliver polished prose. You do need to identify a correction when you change your mind: “Correction: Friday, not Thursday.” If the application cannot preserve that distinction reliably, edit the transcript before asking for a final draft.

Work a short example

Spoken note: “Lee agreed to send the report Friday—actually, said they would check Friday whether it could be sent. We haven't agreed to extend anything.”

Unsafe summary: “Lee will send the report Friday, and an extension is being arranged.” Faithful note: “Lee said they would check on Friday whether the report could be sent. No extension has been agreed.” The difference concerns commitments, not style.

Give the editing stage a boundary

text
Turn this transcript into a concise internal file note. Preserve corrections, uncertainty, commitments, dates, and who said what. Remove filler and repetition. Do not infer an agreement from a discussion. Put unresolved transcription issues in brackets. Finish with actions expressly mentioned, identifying the responsible person only where stated.

For an external email, first decide the audience and the instruction you need. Do not let an internal strategic observation migrate automatically into correspondence. Review the final draft as a communication, including recipients, attachments, dates, and what it commits you to do.

Try it: ambiguous audio

The transcript says “the offer is fifteen/fifty thousand” and the recording is unclear. The model chooses $50,000 because it seems more plausible. What should the workflow do?

Check your understanding

Keep the amount unresolved until checked against the audio or another reliable source. If neither resolves it, mark the ambiguity explicitly. A plausible amount is not a transcription. A useful drafting system should preserve uncertainty instead of hiding it behind professional formatting.

A small quality test

Record a one-minute fictional note containing two names, two dates, a correction, a negative, a conditional commitment, and an amount. Check the transcript before editing, then check the edited note against both. This reveals more about your setup than an impressive demonstration with carefully prepared speech.