Homophone correction

Last updated • 5 min read

Some words sound exactly alike, so a speech model can hear you perfectly and still write the wrong one. Evoglyph fixes context homophones like their versus there and one versus won from the surrounding sentence. It is a skill trained into the cleanup model, bounded by a generated dictionary, and checked before it reaches your text.

The problem: words that sound identical

"Their" and "there." "One" and "won." "Main" and "Maine." These words are acoustically identical, so a speech model transcribes the sound correctly and still has no way to know which spelling you meant. Only the rest of the sentence does: "merged that into main" needs "main," while "drove to Maine" needs the place.

This is the one class of error that listening to the audio again can never resolve, which is exactly where vocabulary boosting stops. Vocabulary boosting is acoustic; it fixes words that were misheard as sound. Homophone correction is semantic; it fixes words that were heard right as sound but are wrong in meaning. The two are deliberate counterparts.

How Evoglyph fixes it

Homophone correction is not a separate model. It is a skill trained directly into the cleanup model. As the model reads your sentence, it proposes the spelling that fits the context. A deterministic check then decides what happens to each proposal: accept it, decline it, or hand it to the safety net to judge.

The accept rule is narrow on purpose. A swap is accepted automatically only when the word being changed and the word replacing it are both listed as alternates of each other in Evoglyph's homophone dictionary. The model can swap a word for a known homophone of that word, and nothing else. Recent refinements widened the safe set to include grammatical homophones such as their, your, and its, and added a guard that declines the trickiest confusable words (whose, their, your, its, than) rather than risk getting them wrong.

Where the dictionary comes from

The homophone dictionary is generated mechanically, not written by hand. It is built from a public pronunciation dictionary, a large word-frequency list, and the built-in macOS lexicon, with automatic filters that drop the words most likely to cause a bad swap: proper nouns, weak pronunciations, and possessives. The result is about 1,868 entries, grouped by how risky a mistake would be:

Every tier is held to a strict "must not touch" standard during evaluation, leaving the correct word alone at least 99% of the time, so ordinary words do not get rewritten out from under you.

What it does, and what it doesn't

Homophone correction runs automatically on every dictation. It handles cases like:

It is not the same thing as two neighbouring features, and it is worth keeping them apart:

The safety net has the final say

The accept-or-decline logic is not the last word. The trickier swaps are escalated to the span-by-span safety net, which can revert a homophone change if it would alter what you said. Homophone swaps are one of the specific edit types the safety net is built to approve or reject, so a correction only survives if it is both a listed homophone and consistent with your meaning.

Known limits

In its most recent full evaluation, homophone correction fixed about 90% of the cases it should fix, up from about 52% in the previous version, while leaving the right words alone about 99.1% of the time. Those figures were measured at that evaluation; the latest refinements (the widened safe set and the new guard) are individually tested but have not yet been re-measured end to end, so read the headline numbers as "measured at the last full run." By design the recent changes should hold or improve those numbers, not lower them.

A small residual of over-correction remains, on the order of one percent. When it happens, the safety net is the backstop, and the worst case stays your own words rather than a wrong swap left standing.

Was this helpful?
Related articles