Is speech recognition more accurate than humans?

Not in general. Machines have reached human level on a few specific tests, mostly clean or telephone speech. On noisy audio, and on speech the system has not learned, careful human listeners still make fewer errors. The honest comparison is always made on the same recording, against the same checked transcript.

One telephone benchmark put machines level with professional transcribers

In 2016, Microsoft researchers reported a system that scored a 5.8 percent word error rate on the Switchboard telephone test ¹. Professional transcribers scored 5.9 percent on the same portion ². The authors called this human parity. On CallHome, a harder test of conversations between friends and family, the system reached 11.0 percent against 11.3 percent for the humans.

A gap of a tenth of a point is a tie for practical purposes. The result is still narrow. It covers one benchmark, one kind of telephone speech and one set of professional transcribers, so it does not show that machines beat humans in general.

The Whisper paper, from 2022, reaches a similar and more cautious conclusion. Its English recognition is "not perfect but very close to human-level accuracy" ³. "Very close" is a weaker claim than "better".

On noisy audio, humans stay far ahead

Noise is where the comparison turns. In a 1997 review of noise experiments, a machine with noise adaptation reached 12.8 percent error at a signal-to-noise ratio of 10 dB, while human listeners made about 1 percent errors at the same level ⁴. Those are old systems, so the numbers show the size of the gap at the time, not today. The lesson still holds: a score on clean audio does not carry over to a café, a car or a windy street.

Human baselines also vary between studies. The table sets the 1997 and 2016 Switchboard results side by side.

On the same telephone test, the two studies' human figures differ, 4 percent in 1997 and 5.9 percent in 2016. The two studies used different transcribers, instructions and reference transcripts. So a "human parity" claim is only as strong as the human baseline behind it.

Four cases where the answer changes

• Clean, read speech: this is the easiest case, and the gap is smallest on modern systems. Even so, a 1997 review found machine error often more than ten times human error on quiet, read speech ⁷, which shows how much the answer has moved since then.
• Live or streamed speech: a 2024 study of eleven services found significantly lower quality for streaming recognition, the mode used for live events ⁸.
• Invented words: a machine can add a fluent phrase that was never said. In a 2024 study, about 1 percent of Whisper transcriptions contained whole phrases or sentences that did not exist in the audio ⁹. A careful human transcriber does not usually do that.
• Names and jargon: a specialist who knows the field often beats a general model on rare terms. The question on names and jargon explains why.

Compare on the same recording

A fair test needs one recording, one reference transcript that someone has checked, and both a machine and a human working on the same audio. Count the errors for each against that reference, using the word error rate formula. A number from a vendor page, or from a study on other audio, tells you little about your own recordings. The human versus AI transcription guide compares cost, speed and accuracy for each option.

What to do

1. Decide what the output is for. A rough draft you will edit can tolerate more errors than a final record.
2. Run the passage test on your own voice and microphone. Read a 200-word passage, dictate it, and count the errors. The accuracy guide explains the steps.
3. Keep a person in the loop when errors are costly. For legal, medical or financial text, a qualified person should check the transcript.
4. Proofread the risky parts. Check names, numbers and the words just after a long pause. These are where a machine error is most likely to look confident.

For the bigger picture, see our overview of speech recognition, and the full guide to speech recognition accuracy. The companion questions cover whether speech recognition is less accurate for some people and why Whisper makes up words nobody said.