Turn a meeting recording into a document you can search

A recording is a poor archive. You cannot search it, you cannot quote from it without scrubbing back and forth, and finding the two minutes that mattered in a ninety-minute meeting means listening to a lot of it again. A transcript fixes all of that.

Recording conditions decide the outcome

Before any settings, understand where accuracy actually comes from. Speech recognition operates on the audio it receives, and no model recovers information that was never captured.

In rough order of impact:

1. Microphone distance. This dominates everything else. A microphone 30 centimetres from the speaker captures dramatically cleaner speech than one two metres away across a table, because the ratio of voice to room noise collapses with distance.
2. Background noise. Air conditioning, traffic, keyboards, and a café are all competing signal. A quiet room is worth more than any processing.
3. Room acoustics. Hard surfaces produce reverberation, which smears the boundaries between words. A carpeted room with soft furnishings performs noticeably better than a glass meeting room.
4. One person at a time. Overlapping speech is genuinely hard, and crosstalk produces the most confusing errors in a transcript.
5. Recording quality. Record at a reasonable sample rate and avoid heavily compressed formats where you have the choice.

Practical version: put the phone in the middle of the table rather than at one end, close the window, and ask people not to talk over each other.

What the transcription produces

The transcription tool runs an English speech model on your device and produces timestamped text. Timestamps are more useful than they first appear: they let you jump back to the original audio at any point, which is what makes verification practical rather than theoretical.

From there you can edit the text and export either a searchable PDF or plain text. The PDF is the right output for filing, circulating, and archiving; plain text is better if the content is heading into another system.

Review before you rely on it

Recognition output is a draft. It will be broadly correct and specifically wrong, and the errors cluster predictably:

• Names. Personal names, company names, and product names have no dictionary support and are frequently mangled. If the meeting was about Acme's Q3 numbers, expect "Acme" to appear three different ways.
• Numbers. Figures, dates, and percentages are where errors cost most.
• Technical vocabulary. Domain jargon is guessed from phonetically similar common words.
• Homophones. There, their, and they're are decided by context the model may read differently from you.
• Speaker attribution. If the transcript does not separate speakers, work out who said what while you still remember.
• Negations. A dropped "not" inverts a sentence, and it is easy to miss when skim-reading.

Read the transcript against the timestamps for any passage you intend to quote, act on, or circulate as a record. For an informal searchable archive, a rougher transcript is perfectly serviceable.

Consent and confidentiality

Two separate obligations, both of which apply regardless of which tool you use.

Consent. The rules on recording conversations vary considerably by jurisdiction — some require only one party to consent, others require everyone. Interviews, medical consultations, and legal discussions typically carry additional requirements. Ask, and note that you asked.

Confidentiality. Recordings of meetings are frequently the most sensitive artefacts an organisation produces, because people say things in conversation that they would never write down. Sending that audio to a cloud transcription service means handing over the unguarded version of a discussion, along with whatever retention and review terms the provider applies.

Local recognition avoids that entirely: the audio is processed on the machine holding it. For interviews, HR discussions, client calls, and clinical notes, that is not a preference — it is often the only defensible option. The same reasoning applies to document assistants, as the AI privacy guide sets out.

After the transcript

A raw transcript is rarely the final artefact. Most of the time what people actually want is a summary with actions and decisions. Edit the transcript into that shape and export it as a document — or compose the summary separately with the document creation tool and keep the full transcript as the underlying record.