Why does speech recognition get names and jargon wrong?

Speech recognition gets names and jargon wrong because those words are rare in the speech it learns from. A model that has seen a common word thousands of times will often choose a familiar word that sounds similar. The errors are measurable: a 2000 study of medical dictation found vocabulary error rates of 14 to 15 percent, and newer systems still show significant errors on rare names. You can reduce the problem, but you have to check the names you use.

Rare names are guesses, because models learn from common words

Recognizers learn from large collections of speech, and common words dominate those collections. Apple researchers report that end-to-end recognition still has a significant error rate for entity names that appear infrequently in its training data ¹. A name the model has rarely seen is a guess among similar-sounding words, and a wrong guess reads like a normal word, so it is easy to miss.

The same pattern appears in legal speech. A 2022 study of UK Supreme Court hearings reports that adapting a generic speech model can avoid critical errors specific to legal jargon and terms used in British courts ². The problem is not limited to one field. It appears wherever the vocabulary is specialist.

The same effect appears in code. A 2026 preprint names custom identifiers, such as variable and function names, as a central difficulty for speech input about code ³. A name you invented has the same problem as a rare drug name: the model has no reason to expect it.

Medical dictation has measured vocabulary errors

Medical speech is the clearest case, and the measurements are older than modern models. In a 2000 peer-reviewed study of physicians dictating medical reports, Dragon NaturallySpeaking Medical Suite 3.0 had a vocabulary recognition error rate of 14.1 to 15.2 percent ⁴. Those versions date from 2000, so the figure shows the scale of the problem, not current performance.

Newer models do better on specialist speech, but 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 ⁵. The same pilot compared it with a commercial medical dictation product. The difference between the two was not statistically significant ⁶. The study had three physicians, so it is a small pilot, not a general result for every specialty.

Overall error rates can also hide the errors that matter. Word error rate counts substitutions, deletions and insertions, and divides by the number of words in the reference ⁷. Every word counts the same, so a wrong dose or drug name costs no more than a wrong "the". A low overall rate can therefore coexist with a wrong name in every note. When you test a tool for medical or legal work, score the key terms separately.

Context and custom vocabulary help, within limits

Some tools let you supply words the model should expect. Whisper's initial prompt can give the model custom vocabulary or proper nouns, which makes those words more likely to be transcribed correctly ⁸. Some dictation tools offer a similar list, often called custom vocabulary. The custom vocabulary term explains how these lists work.

These lists help with terms you can name in advance. They do not help with a term you forgot to add, and they do not fix a recording with heavy noise or unclear pronunciation. The guide on dictation and background noise covers how noise adds errors on top of the vocabulary problem.

What to do

1. List the names and terms before you dictate. Include people, places, products, drug names and abbreviations you use often.
2. Add them to the tool's custom vocabulary, if it has one. Check the spelling and capitalization in the list, since those are what the output will show.
3. Test with a passage that contains your terms. Dictate it, then check each name by hand. The dictation mistakes guide lists the usual error types to look for.
4. Proofread names with extra care. A wrong name looks like a normal word, so skim for it on purpose.
5. Have a qualified person check high-stakes records. For clinical, legal or financial text, a human review is the safeguard, whatever the error rate.

For the wider picture of measured errors, see the guide to speech recognition accuracy. The overview of speech recognition links to the rest of the subject. The companion questions cover whether speech recognition is less accurate for some people, whether it beats humans and why Whisper makes up words.