How accurate is AI transcription?

On clean, single-speaker English, AI transcription is close to careful human work. Whisper Large V2 made 2.7 percent word errors on a clean read-speech test ¹. On noisy, distant or overlapping audio the error rate climbs several times higher. The model can also write whole sentences that nobody said ². Recording conditions change the error rate a great deal: the same model made several times as many errors on the harder test set below.

What a transcription error is, and how it is counted

A transcription error is any place where the written text differs from what was said. There are three kinds: a word is replaced by a different word, a word is left out, or a word is added that was never spoken. Researchers combine these into one figure called word error rate, the proportion of word errors to the words processed ³. Added words count too, so the score can rise above 100 percent in a bad case.

A 2.7 percent score therefore means a few wrong words in every hundred, not a few wrong sentences. It also says nothing about which words, so a wrong name or a flipped number can matter more than the score suggests.

Clean speech is where AI is closest to human level

Clean, read English is the easiest test. The model's authors describe its English recognition as "not perfect but very close to human-level accuracy" ⁴. That is a claim about one kind of audio, read speech, not about a meeting in a café.

Human transcribers are not a fixed yardstick either. In a 2017 Microsoft study, professional human transcribers made 5.9 percent word errors on the Switchboard telephone conversations ⁵. Microsoft's automated system scored 5.8 percent on the same set ⁶. Those numbers are close enough that the two were effectively tied on that test.

Hard audio raises the error rate several times

Accuracy falls quickly when the audio gets harder. On the CHiME6 test set, a harder set with several speakers and background noise, Whisper Large V2 made 25.5 percent word errors ⁷. That is far above its clean-speech rate.

Distance from the microphone matters as much as noise. In the Ravanelli figures from 2017, a single distant microphone produced a word error rate about 3.2 times the close-talking rate ⁸. The two conditions are not identical, so read the ratio as a rough guide.

AI can also invent text, which an average score hides

AI transcription can fail in a way people rarely do. The Whisper authors list complete hallucination, a transcript entirely unrelated to the audio, among its known failure modes ⁹. A few invented sentences add very few words to a long file, so the average error rate can look good while the transcript still contains fiction. Check silences, music and overlapping talk, where invented text tends to appear.

What it depends on

The studies behind these numbers date from 2017 to 2022, and newer models may do better. Read them as the direction of the gap rather than a current ranking. A vendor's accuracy claim also needs a test set and a counting method before it means much, which is why one recording can say little about the next.

What to do

1. Test on your own recording. Pick a few minutes of the audio you actually need, transcribe it, and check it against your own careful reading. That tells you more than any published score.
2. Improve the recording first. A headset or lapel microphone close to the speaker lowers errors, because distance raises them.
3. Check the risky parts. Look at names, numbers, quiet passages and places where two people talk at once.
4. Send only the hard sections to a person. For records that must be exact, a human pass is the safe route. The guide on human versus AI transcription sets out which route fits which recording.

For cost and time, see the answers on how much transcription costs per minute and how long one hour of audio takes. For the wider picture, see our overview of transcription.