Dictation, voice mode and who pays for the errors

The unglamorous technique that saves the most time

Most people type at 30 to 50 words a minute and speak at 120 to 150. If the bottleneck in your work is getting what is in your head into a box, dictation roughly triples the rate, and it is available on every phone and most computers already.

This is not a clever prompting technique. It is the one change that most reliably makes AI useful to people who do not enjoy typing — which, in most workplaces, is nearly everybody.

The division of labour that works

The productive pattern is: you speak the substance, the model organises it.

Walk out of a site visit and talk for four minutes into your phone. Everything you noticed, in whatever order it comes, including the bits you are not sure about. Then paste the transcript in:

This is a dictated site visit note, unstructured and with repetitions. Organise it into: what was inspected, findings, actions required with owners, and anything I flagged as uncertain. Use only what is in the transcript. If I contradicted myself, show both statements rather than choosing.

Notice what has happened. Every fact came from you. The model contributed structure, which is exactly the task where it has no opportunity to invent and no need for knowledge it does not have. This is close to the safest configuration in the whole course, and it is also the one that saves the most time, because four minutes of speech is 500 words you would otherwise not have written.

Accents, names and code-switching

Speech recognition is not uniformly accurate, and it is worth being honest about who bears the cost.

Word error rates vary considerably by accent and dialect, and published work has repeatedly found higher error rates for speakers whose accents are less represented in training data. If you speak English with an Indian, Nigerian, Scottish or Caribbean accent, or if you code-switch mid-sentence between English and Hindi, Yoruba, Tagalog or Arabic — which enormous numbers of people do at work — expect more errors than the marketing demonstration showed, and expect them concentrated in the hardest places to notice.

Three practical mitigations:

• Proper nouns are the weak point. Say an unusual name and then spell it. Do the same for drug names, part numbers and abbreviations.
• Use a custom vocabulary if your tool offers one. Loading fifty terms specific to your trade removes most of the recurring errors in one pass.
• Always check names and numbers against what you meant. These are the fields where an error survives every subsequent step, because nothing downstream can detect that "Mr Ade" should have been "Mr Ede".

Voice conversation mode, and its hidden cost

Speaking to a model and hearing it answer is a different thing from dictation, and the difference matters.

Spoken answers cannot be skimmed. You cannot glance back at the third sentence, notice a figure that looks wrong, and compare it with the second. Rereading is your main error-detection mechanism and audio removes it. A fluent voice is also more persuasive than fluent text — the human ear treats confident delivery as a signal of competence, and here it is a signal of nothing.

So: voice for input, text for anything you have to check. Where a product supports both, dictate your question and read the answer. Where you use voice mode for convenience — walking, driving, hands busy — treat everything you hear as orientation, and confirm anything actionable in writing afterwards.

Accessibility, said plainly

For people with repetitive strain injuries, limited hand mobility, low vision or dyslexia, dictation is not a productivity tweak. It is the difference between being able to do a piece of work and not. Everything in this lesson applies with more force, including the free options below, since accessibility tools have historically been expensive.

The free path, and it is a good one

Whisper, OpenAI's speech recognition model, is released under an open licence. whisper.cpp runs it locally on an ordinary laptop with no network connection at all — a real option when the recording contains a patient, a client or a colleague's performance discussion. Vosk is another offline option, lighter and faster on modest hardware. Both are free.

Local transcription is slower than a cloud service on a laptop — expect a few minutes for an hour of audio on a modern machine, longer on an old one — and for confidential material it is frequently the only version of this that your professional obligations permit. Recording other people brings a separate set of duties, covered in its own lesson later; nothing here changes them.