Dictation in Noise: Speech Recognition Accuracy Tested

Other people talking in the background hurts speech recognition a little in a quiet office, noticeably in a busy cafe, and badly when the talkers are as loud as you are. In our test, all three engines stayed usable until the background talk was about as loud as a busy cafe. Below that level the error rate rose steeply. At equal loudness, none of them produced text you could use without heavy rewriting. Parakeet TDT v3 held up best as the noise grew. Whisper large-v3-turbo was the most accurate in quiet conditions. Apple SpeechTranscriber was the least accurate almost everywhere.

The numbers are word error rates, the share of words that are wrong, missing or added (see speech recognition for how engines work and what the measure means). Lower is better. We ran each engine on the same 100 English recordings, once clean and once at each background talk level, on a single Mac. The table below lists every row, and the chart shows the same figures by level.

Quiet rooms barely matter, and one engine does not notice them at all

Without added noise, Whisper large-v3-turbo made the fewest mistakes at 3.78 percent ¹. Parakeet TDT v3 followed at 4.44 percent ², and Apple SpeechTranscriber at 6.73 percent ³. Those gaps are small in practice. All three are good enough that most sentences come out right.

Add background talk at the level of a quiet office and almost nothing changes for the first two. Parakeet stayed at 4.44 percent ⁴, identical to its clean figure. Whisper moved to 4.20 percent ⁵, which is higher than its clean figure, but the difference is small. Apple moved to 7.61 percent ⁶.

If you dictate in a home office, a library or a quiet open office where voices are faint, background talk is not a reason to pick one engine over another. Pick on other grounds, such as speed, languages and punctuation.

A busy cafe is where the engines start to separate

At the level of a busy office or cafe, the ranking changes. Parakeet barely moved at 4.48 percent ⁷. Whisper rose to 5.51 percent ⁸, so Parakeet overtook it even though Whisper was better in silence. Apple jumped to 13.87 percent ⁹, which is a clear degradation, a rate at which you notice and fix errors constantly.

In a loud cafe, the gap widens. Parakeet reached 7.85 percent ¹⁰, Whisper 13.92 percent ¹¹, and Apple 31.34 percent ¹². Parakeet was still in a usable range. Whisper's error rate had more than doubled in a single step. Apple was making a mistake on a large share of its words.

At equal loudness, every engine fails

When the background talk was as loud as the speaker, the error rates were 44.47 percent for Parakeet ¹³, 66.09 percent for Whisper ¹⁴ and 71.88 percent for Apple ¹⁵. Parakeet still got a majority of the words right. The other two got only a minority right, and their text is closer to a guess than a transcript.

The reason is not hard to see. Babble is made of speech, so it has the same sounds, rhythm and vocabulary as the thing the engine is trying to transcribe. When the voices are equally loud, the engine has no clear way to tell which speech to follow. Ordinary noise such as a fan is easier, because it does not look like words. Following the wrong talker, or dropping stretches of speech, both show up as errors. Our data does not show which of these happened here.

The table shows the three engines at the hardest level:

Each engine degrades differently

Putting the levels side by side shows three different shapes.

• Parakeet TDT v3 is flat from silence to a busy cafe, bends at a loud cafe, and breaks only at equal loudness. It had the lowest error rate at every background level from the busy cafe down.
• Whisper large-v3-turbo is the best in silence and in a quiet office, then loses its lead as the talkers get louder. Its decline is steeper than Parakeet's from the busy cafe onward.
• Apple SpeechTranscriber starts highest in silence and is the first to degrade. It has the highest error rate at every level, including the hardest.

The full set of rows for the engine with the flattest curve:

One caution on ranking: the order in silence is not the order in noise. If you read only a clean-audio benchmark, you would pick Whisper. For a noisy room, the clean figures would mislead you. Our overview of on-device speech recognition benchmarks covers the clean comparison across more engines.

Where studies disagree

This dataset holds one measurement series, our own, so there are no competing studies to set against it. Noise tests from different labs use different noise recordings, loudness definitions and test sentences, so their figures should not be set against ours. What we can say is that the direction of our results is what you would expect: more babble, more errors. The exact positions of the breaking points depend on the setup described below, and we would not expect another lab to reproduce the same figures to the decimal.

What this data cannot tell you

Every row carries the same warning, and it matters. Our background talk is built from read sentences, which is steadier than a real cafe. There are no dishes, no music and no sudden laughter. Real rooms have peaks that catch an engine off guard, so a real cafe is probably harder than the same nominal level in our test. Equally, a close microphone in a real room changes the ratio between your voice and the room, so a level named in the table does not map neatly onto a place. Compare engines and conditions against each other. Do not treat any single figure as what you will get in a real room.

The data also leaves out several things:

• Other languages. Every recording is English. Engines can degrade very differently in other languages.
• Other noise types. Traffic and music are not covered, and they may favor different engines.
• Other hardware and settings. The engines ran on one Mac. A phone microphone, a headset or a different model size could shift the results.
• Whole-document effects. Word error rate counts words, not meaning. A wrong word in the middle of a sentence and a missing "not" count the same, though one is far worse for you.
• Short clips. The recordings are single read sentences, not long dictation with pauses, hesitations and self-corrections.

What this means for choosing how to write

If you write in a quiet place, noise is a non-issue and you can choose any of these engines on speed, privacy and convenience. The clean differences are small.

If you write in a cafe, a coworking space or an open office, choose by the noisy figures, not the clean ones. In our test, Parakeet TDT v3 was the engine that stayed usable the longest as voices got louder. Apple's built-in engine was the weakest option once the room got busy, though it remains convenient and may be enough if you keep the room quiet or use a headset microphone.

If you must dictate where the talkers are as loud as you are, no engine in this test is reliable. The fix is not a better engine but a better signal: move closer to the microphone, use a headset, or find a quieter corner. Every step that raises the ratio of your voice to the room helps more than switching software. And if you do dictate in noise, expect to proofread more carefully, because the errors do not announce themselves.

For a wider comparison of engines in clean conditions, see the on-device ASR benchmark, and for background on how these systems turn audio into text, read our overview of speech recognition.