The real cost per hour of AI transcription, and how to work it out

The advertised per-minute rate is not the cost. The cost is the per-minute rate plus the minutes a human spends fixing what the machine got wrong — and on poor audio that second number can be larger than the first.

Transcription is one of the few AI categories where the real economics are easy to calculate, and almost nobody calculates them. Vendors quote a rate per minute or per hour of audio. Buyers compare those rates. The comparison is nearly meaningless, because it prices only the part of the work a machine does reliably.

The model

For any transcription tool, the true cost per hour of audio is:

Machine cost — the per-minute rate multiplied by sixty, times one plus the rework rate. The rework rate is the share of audio you end up paying for twice because the first pass failed and you reran it.

Review cost — the number of minutes a human spends per hour of audio checking and correcting, multiplied by their loaded hourly cost, divided by sixty.

Platform cost — any monthly subscription, divided by the hours of audio you actually process each month. This is where small users get destroyed: a subscription you use twice a month can cost more per hour than the audio itself.

Add the three. That is your real number.

Worked example

Take a team processing twenty hours of audio a month, on a tool priced at a rate typical of the mid-market, with a subscription on top, and a reviewer on a loaded cost of forty currency units an hour.

• Machine cost: the per-minute rate, times sixty, is the headline figure the vendor shows you.
• Rework: assume one hour in twenty needs a rerun because of overlapping speakers or a bad connection. That is a 5 per cent multiplier.
• Review: the honest figure for a clean two-person interview is roughly ten to fifteen minutes of checking per hour of audio. For a six-person meeting with crosstalk, twenty-five to forty minutes is realistic.
• Platform: divide the monthly subscription by twenty hours.

The result worth noticing is the review line. At fifteen minutes per hour, a reviewer costing forty an hour adds ten units to the cost of every hour of audio — often several times the machine cost. Your transcription bill is a labour bill with a software tip attached.

Where the model changes the decision

The arithmetic produces three conclusions that the headline rate hides.

Accuracy is a cost multiplier, not a feature. A tool that is five per cent more accurate but twice the price is usually the cheaper option, because accuracy moves the review line and price does not.

Volume changes which plan wins. Subscription pricing favours heavy users and punishes light ones. Below roughly ten hours a month, per-minute billing almost always beats a subscription once you divide it out.

Meeting structure beats tool choice. A recording with one speaker at a time transcribes far more accurately than a room of six people interrupting each other. Moving from an open room to a round-robin saves more money than switching vendors.

What we don't know

Review time per hour is the single most important input and the one with the least public data. Published accuracy figures are usually measured on clean, read speech — studio recordings and audiobooks — which is not what your Tuesday client call sounds like.

We also cannot tell you the rework rate for your own audio without measuring it, and vendors do not publish failure rates by audio condition.

What to do next

1. Take ten hours of your own recordings and measure the review minutes per hour. Do it once, properly, with a timer. That number will decide more than any vendor comparison.
2. Divide your monthly subscription by your actual monthly hours. If the result embarrasses you, move to usage billing.
3. Switch one recurring meeting to a round-robin format — one speaker at a time, acknowledged handovers — and re-measure. If the review time drops, you have found a saving that costs nothing.