Does dictation work in a noisy room?

Dictation works in a moderately noisy room and breaks down in a very loud one. In our own test of three on-device engines, accuracy barely moved while background talk stayed well below the speaker's voice. It fell apart once the other voices became as loud as the person dictating. The engines differ a lot in where that happens, so the choice of engine matters more in noise than in a quiet room.

Quiet background talk costs almost nothing

With background talk 20 dB quieter than the speaker (about a quiet office), two of the three engines were nearly unchanged. Parakeet TDT v3 made errors on 4.44 percent of words ¹, the same as its figure with no noise at all, 4.44 percent ². Whisper large-v3-turbo moved from 3.78 percent ³ with no noise to 4.20 percent ⁴. Apple SpeechTranscriber went from 6.73 percent ⁵ to 7.61 percent ⁶.

If you dictate in a quiet office, a coworker talking across the room is not the thing to worry about.

A busy office or cafe separates the engines

At 10 dB, which stands for a busy office or cafe, the gap opens. Parakeet TDT v3 stayed at 4.48 percent ⁷. Whisper large-v3-turbo rose to 5.51 percent ⁸. Apple SpeechTranscriber rose to 13.87 percent ⁹, well above its no-noise figure.

Every engine here is still usable at this level, but the errors are now frequent enough that you will notice them when you proofread.

A loud cafe is where Apple's engine falls off first

At 5 dB, a loud cafe, Parakeet TDT v3 reached 7.85 percent ¹⁰. Whisper large-v3-turbo reached 13.92 percent ¹¹. Apple SpeechTranscriber reached 31.34 percent ¹², so a large share of the words were wrong.

This is the level where the three engines look least alike. The ranking was the same at 10 dB and every louder level we tested: Parakeet first, Whisper second, Apple third.

When voices are as loud as yours, none of them cope

At 0 dB the background talk is as loud as the speaker. Parakeet TDT v3 had a word error rate of 44.47 percent ¹³. Whisper large-v3-turbo had 66.09 percent ¹⁴. Apple SpeechTranscriber had 71.88 percent ¹⁵.

At those rates the output is a rough guide at best. An error rate this high usually means the engine is also transcribing the wrong voice, not only misreading yours. Dictation is not practical here with any of the three.

The numbers by condition

The tables below list every engine at the two ends of the range that matter most for choosing a tool.

The full set holds 15 data points ¹⁶: three engines, each measured once with no noise and at four background talk levels. You can browse all of them in the full dataset at /data/asr-noise-robustness.

What these numbers do and do not measure

The test used 100 English recordings of read sentences from the FLEURS collection, run on a Mac with an M2 Pro in October 2026. We mixed other people's speech under each recording at a fixed level below the speaker. Word error rate is the share of words that are wrong, missing or extra, so lower is better.

Four limits apply:

• The noise is steadier than a real cafe. Babble built from read sentences has no dishes, music or sudden laughter. Real rooms are messier.
• The microphone is idealized. The speaker and the babble were mixed digitally. A microphone close to your mouth in a real room changes the ratio between your voice and the background, usually in your favor.
• English only. Other languages and other noise types such as traffic or music were not part of this test.
• One setup per engine. Results can shift with a different Mac, a different version of an engine or different settings.

Compare engines and conditions with each other. Do not read any single figure as the error rate you will get in a specific room.

What this means for how you dictate

First, the biggest single fix is the ratio between your voice and the room, not the software. A headset or a microphone close to your mouth makes your voice louder relative to everyone else, which moves you toward the left of the chart where all three engines do well.

Second, if you often dictate in a busy office or cafe, pick the engine by its noise results, not only its clean-room score. In our test the engine with the lowest error rate at 0 dB also changed least between no noise and a busy room. The engine that was weakest with background talk was the weakest in quiet too, but its drop in noise was much steeper. For a side by side of products, see /market.

Third, if the room is as loud as you are, move or pause. No engine in this test rescued speech at that level.

Where to read more

For how accuracy is measured and what drives it, read our guide to speech recognition accuracy and the overview of speech recognition. For a quiet-room comparison of on-device engines, see the on-device ASR benchmark.