Brief Buck
Start a campaign
✦ Task type · Language work

Transcription, now that the machine does the first pass

Automatic transcription is accurate enough that paying for a first pass by hand is usually waste. What it still cannot do is know which words matter.

Post this taskThinking of taking one instead?

Transcription is the clearest example of a task type that automation changed rather than removed. The machine handles the ninety per cent that was always mechanical. What is left is judgement about the remaining ten.

What to put in the public brief

Keep the audio itself private. It is the asset.

What to put in the private instructions

  1. Give the audio, the machine draft, and the format

    A named format — speaker labels, timestamp interval, whether to keep filler words — removes every decision they would otherwise make differently from you.

  2. List the proper nouns in advance

    Names, products, places, internal jargon. This single step removes most of the errors that survive machine transcription, and it takes you two minutes.

  3. Say what to do with inaudible passages

    A marked timestamp and a flag, never a guess. A plausible invented word is worse than a gap, because a gap is visible and a guess is not.

  4. Say whether to clean it up

    False starts and repetitions kept or removed. Both are correct for different purposes, and neither is guessable.

Proof worth requiring

Ask for

The transcript in the specified format, plus a list of the timestamps they marked inaudible. That list tells you where to listen yourself.

Do not ask for

A confidence score or an accuracy percentage. Nobody can estimate their own accuracy on their own work, and the number will be optimistic in a way that is not dishonest, only human.

When this task type is the wrong tool

If the audio is legal, medical or otherwise regulated, use a service with confidentiality obligations and liability, not a marketplace.

If you need it in another language as well as transcribed, that is two tasks. Short-copy translation covers what changes when the output has to work in a different market rather than only a different alphabet.

Common questions

Should I pay for transcription or correction?

Correction, almost always. Run the audio through automatic transcription first and pay a human to fix it. You get better accuracy for less money, and the errors that remain are concentrated exactly where a human was needed: names, jargon, crosstalk and anything said quietly.

What accuracy standard should I ask for?

Say what the transcript is for, then set the standard from that. Verbatim including false starts for research; clean read for publication; timestamps every thirty seconds for video. Asking for "accurate" gets you somebody's guess at which of those you meant.

What reward is fair?

Around one to two dollars per audio minute for correction, two to four for transcription from nothing. Multiply for poor recording quality, heavy accents or more than two speakers, because those change the work far more than duration does.

Is my audio confidential?

Uploaded files are stored privately and are not part of the public brief. That said, decide before publishing whether the recording contains anything that should not be heard by someone outside your organisation, because a transcription task means a stranger listens to all of it.

Thinking of taking this task?

The same job, from the other side. Rewards are set by the poster, so treat these as the range this task type usually sits in rather than a promise.

Usually pays
$1 to $4 per audio minute
You need
Accurate hearing; headphones help
Proof expected
The transcript in the format named

Usually you are correcting machine output rather than typing from nothing. Mark anything inaudible with a timestamp instead of guessing, because an invented word is worse than a gap.

Create a worker account

Brief a transcription

Set the reward, the audio length and any accent or language requirement. Workers reserve a spot; your audio stays private until they do.

Post this task