Recording, transcription and the consent question

Ask before you press record

Start with the part that has consequences, because it is the part people skip.

In much of the world, recording a conversation requires the agreement of the people in it. In the United States the rule varies by state: some allow recording with one party's consent, while others — California, Florida, Illinois, Pennsylvania and several more — require everyone's. Under the GDPR in Europe and comparable regimes elsewhere, a recording of an identifiable person is personal data, which brings duties about purpose, retention and access whatever the meeting was about.

Beyond law there is the practical fact that people speak differently when recorded, and a colleague who discovers afterwards that a difficult conversation was transcribed by a third-party service will remember it for years.

So: announce it, in the invitation and again at the start. Say what tool, what happens to the recording, and how long it is kept. Offer to stop. For anything sensitive — grievances, health, performance, redundancy — the answer is usually not to record at all, and to take notes like a professional.

If the tool is not one your employer has approved, this is also the moment to remember that you are sending a recording of your colleagues to an outside company. That is the subject of the next module and the rule there applies here first.

What machine transcription gets wrong

It is much better than it was and it fails in ways that surprise people.

It invents. Whisper, the open-source model behind a great many transcription products, has been shown by researchers to produce entire fluent sentences that nobody spoke — most often during silences, background noise, or the pauses of a hesitant speaker. Not garbled text: clean, grammatical, plausible sentences. That is the same mechanism as everything else in this course, arriving in an unexpected place.

It is uneven across speakers. Accuracy is markedly worse for strong regional accents, second-language speakers, older voices and anyone speaking quietly or over others. If your meeting has a confident native speaker and a quiet colleague on a poor connection, the transcript over-represents one of them. That is not a neutral technical limitation; it decides whose contribution ends up in the record.

It attributes badly. Speaker labels — diarisation — are the least reliable part of the output, especially on a shared microphone or when people interrupt. "Speaker 2 agreed to the deadline" is exactly the sentence you must not trust without listening.

It cannot spell your world. Names, drug names, product codes, local place names and technical terms come out as the nearest common word. A supplier called Bhaskar becomes "Bashkar" or "basket".

Getting a usable result

• Fix the audio, not the model. One microphone close to the speakers beats any amount of post-processing. Ten seconds spent moving a laptop is worth more than a paid upgrade.
• Give it a vocabulary list. Most tools accept a list of names and terms. Adding the eight people in the meeting and your product names removes most of the embarrassing errors.
• Never circulate a raw transcript. It contains errors, it contains half-sentences, and it makes everyone look inarticulate. Work from it; do not publish it.
• Do the summarising separately. Bundled meeting summaries are usually weak. Take the transcript into a chat tool and ask for exactly what you need: decisions, owners, deadlines, and open questions — with the speaker and the approximate time for each, so every line can be checked.
• Check at the timestamps that matter. Not the whole recording. The three or four lines that commit somebody to something. Listen to those.

The free path

Whisper is open source and free, and runs on an ordinary laptop through whisper.cpp or one of several desktop wrappers, with no audio leaving your machine — which makes it the right choice for anything confidential as well as the cheapest. On a phone, Google's Recorder and Live Transcribe cost nothing. Paid services buy you convenience, calendar integration and better speaker labelling, not fundamentally better words.

The thing worth protecting

There is a version of meeting AI that quietly changes how a workplace behaves: everything recorded, everything searchable, every half-formed idea preserved and attributable forever. People notice, and they stop thinking out loud.

The half-formed idea is where most good work starts. A team that will not risk one has lost something no transcript recovers. Record the meetings that need a record. Leave the rest alone.