Human transcription, AI transcription, and the difference that matters

What automatic transcription does well

Automatic speech recognition has improved enormously, and it is worth being precise about where. Given a single speaker, a good microphone, a quiet room, a familiar accent and vocabulary in ordinary use, modern recognition is fast and largely accurate. For a personal note, a rough search index, or a first look at what is in a file, it is entirely adequate.

It is also the reason our lowest rate exists. Where audio is clean, letting a machine produce the first draft saves time that would otherwise go on typing, and that saving is passed on to you at $0.89 per audio minute.

Dismissing the technology would be dishonest. The question is not whether it works. It is what happens when it does not.

Where it fails, and how the failure looks

The failure mode is the important part. Automatic transcription does not stop when it becomes uncertain. It produces a fluent, grammatical, plausible sentence that is wrong.

Human speech breaks recognition in predictable ways. Two people talking at once. A name or a technical term the system has never encountered. A regional accent. A word that sounds almost identical to another word with a different meaning. In each case the system commits to its best guess and presents it in exactly the same tone as everything else.

Nothing in the output marks the difference. A passage the system effectively invented looks identical to a passage it transcribed correctly. There is no flag, no visible uncertainty, nothing to tell a reader which sentences deserve a second look.

A human transcriptionist behaves differently under the same conditions. Where a passage genuinely cannot be made out, it is marked as inaudible with a timecode. The uncertainty is recorded rather than papered over.

Why a confident error is worse than a gap

A marked gap is a small, contained problem. Someone can go to the timecode, listen, and resolve it. It costs a minute, and it is visible.

A confident error is not contained. It reads naturally, so nobody checks it, and it travels into whatever the transcript is used for. Into the quotation in the article. Into the summary sent to the client. Into the evidence bundle, the research dataset, the compliance file.

The damage is not proportional to the number of wrong words. A transcript can be highly accurate overall and still be unusable, if what it got wrong was the name of the drug, the figure in the settlement, or which of two speakers made the admission.

Where we use AI, and where we do not

Our position is a single rule, and it has no exceptions. AI may produce a first draft. Review and quality assurance are human, on every transcript, on every tier.

On the $0.89 tier a machine produces the initial draft, and our transcriptionists then review it against the audio across multiple passes, correcting as they go. The document you receive is human work, regardless of how the first version was generated.

We do not use AI to check AI. A second model reviewing the first one's output shares its blind spots, and adds a second layer of confident error on top of the first. Verification only means anything when it comes from outside the system that produced the text.

Choosing between the tiers

If your audio is clean and budget matters, the AI-drafted tier at $0.89 is the sensible choice. The review is identical, and on straightforward recordings the draft has little in it to correct.

Where the material is difficult, sensitive, or will be relied on by someone else, human from first pass at $1.19 is worth the difference. A transcriptionist typing from the audio hears the whole recording from the start, rather than reading a machine's interpretation and hunting for mistakes inside it. Those are different tasks, and the second is more prone to accepting a plausible error.

Both tiers carry the same 99.00% accuracy target and the same multi-pass human review. If you are unsure which suits a particular file, send us the details and we will tell you which we would use.