How do you improve dictation accuracy?

Before you change the words you choose, check the sound the engine receives. Start with the microphone distance and the background noise, because they change that sound. Then speak in complete sentences so the engine can use context to choose between similar words, add names to a word list where your app allows it, and proofread by meaning. This answer covers each step, the evidence for it, and the cases where it does not help. The full set of causes is in the dictation mistakes guide, within the dictation topic.

Move the microphone closer and quiet the room first

Distance is the first factor to check. A study that re-recorded read speech with a distant microphone found that recognition performance was dramatically reduced compared with close-talking speech ¹. The close-talking baseline in that study was recorded in a studio, and it had a word error rate of 3.7 percent with DNN models ². That is a best case for read speech, so everyday dictation is likely to make more errors. The direction is clear, though: a closer microphone does better.

Nuance, which sells the Dragon dictation software, gives distance guidance for its products. For a handheld microphone it advises keeping the microphone element 1 to 3 inches from the mouth ³. Treat that as a starting point. The right position depends on how you hold the microphone, and a fixed desk microphone across the room is a different situation.

Noise matters too. In a 2004 study of an oral-history archive with noisy, elderly and accented speech, signal-to-noise ratio and syllable rate were the two dominant factors in the overall error rate ⁴. That study used a 2004 recognizer, so its absolute error rates do not describe today's engines. The ranking of the factors is the useful part. Close the window, turn off the fan, and move somewhere quieter before you change your vocabulary or your phrasing.

Speak in full sentences so the engine has context

A speech recognizer chooses between sound-alike words using the words around them, so full sentences give it more to work with. Nuance's Dragon guidelines recommend dictating in longer phrases rather than short phrases or single words, because hearing a whole phrase helps the software predict what you say ⁵. That is a vendor's guidance for its own product. We did not find an independent measurement that compares phrase-by-phrase and word-by-word speech on the same material, so test it on your own draft. Dictate one paragraph in fragments and another in full sentences, then count the errors in each.

Add names and jargon where the app allows it

Names and technical terms repeat, so fixing them once saves time. Dragon's help describes adding a word by specifying the word and its spoken form, and optionally training the software on its pronunciation ⁶. Cloud speech services offer a similar weighting feature, but its effect is limited by default. Google says that model adaptation by default has a relatively small effect on recognition, especially for one-word phrases ⁷. If your app has no word list, the method in how to make dictation spell names correctly is a more dependable plan than waiting for the engine to learn.

It depends on the speaker and the engine

Accuracy is not the same for every speaker or every engine, so a technique that works for one person may not close the gap for another. In a 2020 study of five commercial speech recognition systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers ⁸. That study used interview audio and systems from 2020, so it does not predict your results. It does show that an engine can perform very differently for different voices. If the same words fail every time after you have fixed the microphone and the room, try a different engine or app before you rewrite everything you say.

Check the filters before you blame the engine

Some changed words are not recognition errors. A profanity filter or number formatting replaces words after the engine has produced them. If a word looks altered on purpose, check the setting first. See why talk to text censors words.

What to do, in order

1. Position the microphone at the distance your device or product recommends, and keep it there while you dictate.
2. Quiet the room. Close the window, turn off nearby fans, and move away from the TV or traffic.
3. Dictate whole sentences and pause at the end of each one, not in the middle.
4. Add names and jargon to the word list if your app has one, or fix them in review.
5. Proofread by meaning using the method in how to proofread dictated text.
6. Measure the change. Dictate the same paragraph before and after a change, and count the errors in each.

The steps that change the sound come first, because a better sentence cannot recover a sound that noise has covered. For the room, see dictation and background noise. For the hardware, see the best microphone for dictation.