How Accurate Is Speech Recognition? The Real Numbers

Accuracy depends more on the audio than on the brand. On clean, read English, the best result the Whisper paper cites, by 2021, misses about one word in every hundred ¹. The same kind of system misses about 14 words in a hundred on telephone conversation ², and about 25 on CHiME6, a different test set in the same paper ³. Every figure here is a word error rate, or WER: the substituted, deleted and inserted words, counted against a reference transcript and divided by the number of words in that transcript ⁴.

The useful question is not "how accurate is speech recognition?" but "how accurate is it on audio like mine?" The sections below give measured numbers for each condition, then a ten-minute test to measure your own.

On clean, read English, the best recognizers miss only one to three words in a hundred

Clean read speech is the easiest common test. The LibriSpeech test-clean split is the benchmark most often quoted for it. The Whisper authors report that the best LibriSpeech test-clean error fell from 5.3 percent in 2015 to 1.4 percent by 2021 ¹. Their own large model scored 2.7 percent on the same clean set ⁵.

A percentage is easier to judge in words. At a rate of about one in a hundred, a short sentence usually comes out right, while a long report at the same rate still contains several errors. The errors also differ in kind. A substitution swaps one word for another, a deletion drops a word, and an insertion adds a word that was never said. An insertion or substitution that reverses a meaning, such as "not" appearing or disappearing, matters more than the same count of harmless errors.

Those are benchmark results on clean recordings. The table puts them next to harder conditions. The rows use different test sets, so read it as a map of difficulty, not as a ranking of products.

The medical row is the exception to the pattern. A dictation model used in one specialty can be accurate even on technical speech, as the section on names and jargon explains.

If you dictate in a noisy or conversational setting, the clean-speech number is the best case, not the expected one. The guide to how Whisper works in practice covers the model itself.

Real conversations, background noise and distant microphones still multiply the error rate

Noise is the clearest case where clean results stop applying. In a 1997 review, Richard Lippmann compared machines with human listeners. In one controlled test, a machine with noise adaptation reached a 12.8 percent error rate at a signal-to-noise ratio of 10 dB. That was more than ten times the 1 percent human error rate ⁸. Steady noise at one level is a simpler case than a café or a street, so real-world errors can be larger.

Conversation is the hardest common case. A 1997 HMM recognizer in the same review made 43 percent errors on Switchboard, a corpus of spontaneous telephone conversations ⁹. By 2017, a Microsoft system reached 5.8 percent on Switchboard and 11.0 percent on CallHome, where friends and family talk ¹⁰. Conversation is still harder than read speech: the 2022 Whisper model made 13.8 percent errors on Switchboard ², several times its error on clean read English.

Part of the reason is how people talk. Speakers overlap, restart sentences, swallow word endings and fill pauses with hesitations. The Open ASR Leaderboard accounts for this: before it scores a normalized transcript, it removes disfluencies such as fillers, repetitions and repairs ¹¹. Those are the very features a dictation tool has to handle when you speak naturally.

Live, streaming recognition adds another penalty. A 2024 study of speech recognition services found significantly lower quality for streaming recognition, the mode used for live events ¹². For the latency side of that trade-off, see the guide to real-time speech recognition.

Leaderboard averages hide this spread, too. The Open ASR Leaderboard, covered in the guide to speech recognition benchmarks, computes its averages after punctuation, casing and fillers are removed ¹¹, so they will not match the raw text a dictation app shows you.

Distance from the microphone can roughly triple the error rate

In real recordings made in an apartment with a single distant ceiling microphone, a DNN recognizer had a word error rate of 12.0 percent ¹³. Against close-talking studio speech, the distant rate was about 3.2 times higher ¹⁴. The two conditions are not identical, so treat the ratio as a guide to the size of the effect, not a constant.

A fair test of distance changes one thing at a time. Play the same speech through the same room, score it against the same reference transcript, and move only the microphone, for example from the desk to the sofa. Any difference you see then belongs to that one change.

The practical lesson is the one the dictation guides already give: keep the microphone close and the room quiet. Dictation and background noise covers where to dictate, and the microphone guide covers what to buy.

Accuracy differs across accents and demographic groups, and studies have measured the gap

One measured gap by race comes from a 2020 study of five commercial systems. The average word error rate was 0.35 for Black speakers and 0.19 for white speakers ¹⁵. Those are the study's averages across five systems, not the error of any one product.

The gap has a simple cause. Recognizers learn the sounds and words that dominate their training data. Speech that differs from that data, whether by accent, age, a speech impairment or a dialect, starts further from what the model expects. The fixes in the studies below work the same way: they add examples from the speakers who are poorly served, and the error rate falls.

Atypical speech shows bigger gaps. In Google's Project Euphonia study of 432 people with disordered speech, personalized models reached a median word error rate of 4.6 percent, against 31 percent for general speaker-independent models ¹⁶. Parkinson's disease shows the same pattern. A baseline recognizer had a 3.4 percent error rate on typical speakers and 36.3 percent on speech from people with Parkinson's, and fine-tuning on data from the Speech Accessibility Project brought the second figure to 23.7 percent ¹⁷. The guide to speech recognition for speech impairments covers what those fixes look like in an app.

Children are harder again. A 2025 study found that a child's age and the amount of speech in the audio had the largest effect on accuracy, followed by background noise and pronunciation ability ¹⁸.

Not every study finds an accent effect. A 2004 study of the MALACH oral-history archive found no evidence that accent or speaker age affected recognition ¹⁹. Accent is one factor among several, and the audio often matters more. For how accent affects results in particular, read the guide on speech recognition and accents. For the question itself, see whether speech recognition is less accurate for some people.

Rare names and specialist jargon are a weak spot for recognizers

Recognizers learn from common words. Apple researchers report that end-to-end recognition still has a significant error rate for entity names that appear rarely in its training data ²⁰. A name the model has rarely seen is a guess. A single wrong name costs one word in the error count, but it can cost much more in trust, because a reader notices it at once. Check every name and product term once in a test passage, and you learn which ones the tool gets wrong.

Older medical dictation shows the pattern clearly. In a 2000 peer-reviewed comparison of packages used to write medical reports, Dragon NaturallySpeaking Medical Suite 3.0 had a vocabulary recognition error rate of 14.1 to 15.2 percent ²¹. That software is old, but the lesson about specialist vocabulary still holds.

Modern models do better on specialist speech, though the evidence is thin. In a 2025 neurosurgery pilot of a Whisper-based dictation model, the mean word error rate was 1.75 percent, counting all errors ⁷. A small pilot shows the result in that setting. It does not show the same for every specialty. The guide on dictation mistakes lists the common error types and what to check for each.

Whisper-type models can invent fluent text nobody said, a different failure from mishearing

Ordinary misrecognition substitutes or drops words that were spoken. Some neural models also produce text that was never spoken. The Whisper authors list "complete hallucination" among known failure modes, where the model outputs a transcript entirely unrelated to the actual audio, alongside repeat loops and missing words at the edges of a segment ²². The authors treat these as known limitations, not as a measured rate.

The rate matters. A 2024 study of Whisper transcriptions found that roughly 1 percent contained entire hallucinated phrases or sentences that did not exist in the audio ²³. That study tested the 2023 version of the model ²⁴, so newer releases may differ.

Silence and background sound are a known trigger. A 2025 study played non-speech sounds to Whisper and found a set of hallucinations that appear frequently ²⁵. Those are artificial test inputs, so the study shows the failure mode, not how often you will meet it. The practical check is simple: after a long pause, look for text that does not match what you said. The question on why Whisper makes up words explains the mechanism, and the Whisper guide explains which tools use the model.

The safe habit is to proofread dictated text the way you would check a colleague's draft. Give most attention to passages that follow a long pause, and be cautious with any recording that is mostly silence or background noise, where a hallucinated phrase is most likely to appear.

Professional human transcribers make errors too, so human parity is a moving target

The claim that a machine matches a human is only as strong as the human baseline. In the 2017 Microsoft study, professional transcribers had a 5.9 percent word error rate on the Switchboard portion of the NIST 2000 test and 11.3 percent on CallHome ²⁶. The automated system reached 5.8 and 11.0 percent ¹⁰, which the authors reported as human parity on that test. The result holds for that benchmark and that transcription setup, not for every human and every recording.

An earlier human figure is lower still. Lippmann's 1997 review put the human error rate on Switchboard at 4 percent, more than an order of magnitude below the 43 percent machine error in the same review ²⁷. The human figures come from different tests, so compare machine with human within each study, as the table and chart below do.

A fair human baseline needs the same audio, the same instructions and a reference transcript that has itself been checked. A human figure is an estimate: it depends on who transcribed, how carefully, and how the reference was made. That is why a claim of parity is always a claim about one test.

The human versus AI transcription guide compares accuracy, cost and speed for each option. The question on whether speech recognition beats humans takes the comparison further.

Error rates on telephone speech fell about sevenfold between 1997 and 2017

Comparing the two Switchboard results gives a rough measure of progress. A 1997 HMM recognizer made 43 percent errors ⁹, and the 2017 Microsoft system made 5.8 percent ¹⁰. That is about 7.4 times fewer errors ²⁸. The two systems use different methods and slightly different tests, so the ratio shows the scale of the change, not an exact rate of improvement.

Progress on the cleanest benchmark continued after 2016. The best LibriSpeech test-clean error went from 5.3 percent in 2015 to 1.4 percent in 2021 ¹. Those gains are measured on clean, well-recorded audio. Benchmark numbers also leave out what a user actually experiences: the time spent correcting mistakes afterwards, and the names and terms the tool has never met. Neither shows up in a word error rate. The history of speech recognition puts these numbers in their longer context.

You can measure accuracy on your own speech in ten minutes with a known passage

A published figure tells you how a system did on someone else's audio. Your own number is the one that decides whether a tool works for you. The test below needs a text editor and about ten minutes, and it uses the same error count that researchers use.

1. Pick a passage of about 200 words that you can read aloud, and type it into a file. That file is your reference transcript.
2. Read the passage aloud in the room and with the microphone you normally dictate with, at your usual pace.
3. Run the recording through the tool you want to test, and save its output. If you prefer a browser, free online dictation works for a quick check.
4. Lowercase both texts and remove punctuation before counting, as the Open ASR Leaderboard does ¹¹.
5. Count the substituted (S), deleted (D) and inserted (I) words in the output against the reference, and divide by the number of words in the reference (N) ⁴: WER = (S + D + I) / N. A word-alignment tool does the counting if you have many passages.
6. Repeat with two or three passages and in two or three rooms, because one reading is a sample.

One test is a sample. A 2025 test of a single seven-minute, 31-second episode gave Apple's speech API a word error rate of 10.3 percent ²⁹, and gave Whisper Large V3 Turbo 1.5 percent ³⁰. The author notes that the exact numbers would differ under slightly different methods ³¹. Treat any single figure, including your own first test, as a rough guide and repeat the test before you decide. Keep the passage and the scores. Run the same test again after an update or a change of microphone, and the comparison tells you whether the change helped your audio, which is the only comparison that matters for your work.

Yap, which we make, runs dictation on the device on Mac and Windows, so the same passage test applies to it. You can download it here and run the test in your own setting. For how the error count is defined and read, see the word error rate guide.