What does it mean to transcribe audio?

To transcribe audio means to listen to a recording and write down what was said, word for word. The written result is called a transcript. Software can now produce a first draft from a clear recording quickly, and a person then checks that draft against the sound. The draft is only as reliable as that check.

Transcribing means writing down the words as they were spoken

A dictionary entry gives the everyday meaning. If someone asks you to transcribe something, they want you to listen to it and write down what was said, word for word ¹. Speeches, interviews and trials are common examples, because the written record is what people need later.

The word covers two jobs. A person types while listening, or a machine does the listening and produces the text. Many transcripts today start as machine output. A speech recognition system turns the sound into text, and then a person corrects it. A professional who does this work is a transcriptionist.

Accuracy is measured by word error rate

One common measure of a transcript is its word error rate. Researchers describe word error rate as a standard measure of the discrepancy between machine and human transcriptions ². It counts the words the machine got wrong, as a share of the correct text ( word error rate explained).

On a clean recording, the rate can be low. Whisper large-v2, an open speech recognition model, had a word error rate of 4.2 percent on English in the Fleurs benchmark ³. That is about four errors in every hundred words. It makes a useful draft, but it is not a finished record, and the rate rises as audio gets harder.

Some computers already transcribe recordings for free

Transcription is not limited to paid services. Apple's Voice Memos app on a Mac can recognize and transcribe the speech in audio recordings ⁴. Open models such as Whisper, and paid cloud services, cover longer files, more languages and more control over the output. Our overview of how to transcribe audio for free lists the free options.

Dictation is related but different. Dictation turns your own speech into text as you talk, so no recording is involved. Dictate or transcribe? explains the difference in more detail.

Proofreading is what makes a transcript reliable

Treat the machine output as a draft. Play the recording at normal speed, stop at each sentence that does not match the sound, and correct it. Names, numbers and dates need the most care, because one wrong digit can change the meaning. Then check the quiet parts, where a tool may have added text that was never spoken.

How much checking you need depends on the use:

• Personal notes: one careful read against the sound is usually enough.
• Clear speech with one speaker: an AI draft plus a proofread is a sound method.
• Formal records, legal or medical use: a full human review is the safer route, though human transcription also makes errors.

Even the best recent recognizers are not flawless. The Whisper authors concluded that its English recognition is very close to human-level accuracy, though not perfect ⁵.

What to do

1. Decide what the transcript is for: notes, subtitles, a quotation or a legal record.
2. Get a draft from a tool you trust with the recording's privacy level.
3. Proofread the draft against the audio, checking names, numbers and dates first.
4. Choose a style. Some uses need every filler word, while others need a clean version. The guide to verbatim and clean verbatim explains the choice.

For the full method, see our guide to transcribing audio to text, which is part of our overview of transcription. Other questions cover long recordings, audio from a link and audio playing on your computer.