Voxor
Download 
Say “Ask Wojciech about the ClickHouse migration” into a general-purpose dictation app and there is a fair chance you will get neither name right. You fix the sentence, send the message, and move on. Tomorrow, the app makes the same mistake.
This is one of the most irritating failures in dictation. The recognizer can handle the whole sentence, yet the two words that carry the most information are the ones you have to repair.
Most dictation tools offer a custom dictionary for exactly this reason. The feature makes sense. The chore does not. You have to stop what you are doing, open settings, remember which words have been causing trouble, type each one, and sometimes add a pronunciation. Almost nobody keeps that list current.
We wanted the dictionary in Voxor to grow during normal work instead.
The words change with the work
A personal vocabulary is rarely a collection of permanently obscure words. It follows whatever has your attention this month.
A new project brings repository names, package names, customer names, and a fresh set of acronyms. Move to another team and half of those terms disappear from daily conversation. Even familiar words can become unusual in context: Rust, Go, Swift, Linear, and Cursor all mean something specific around a developer’s desk.
Importing contacts or scanning recent documents can help. It gives the recognizer a shortlist of words it may encounter. It cannot tell us which names you actually say, or whether “linear” in the current sentence is a product or an adjective.
The best clue usually arrives when you fix the transcript yourself.
Learn from the correction that already happened
Imagine Voxor hears vector and you change it to Voxor. That tiny edit is
more useful than a long setup form. We have the original sound, the spelling you
wanted, and the sentence in which you wanted it.
If the same replacement happens again, confidence grows quickly. If Voxor
also appears in recent documents, the evidence becomes stronger. A manual
dictionary entry is stronger still because it is an explicit instruction.
Those clues do not all deserve the same weight. A name found once in an old document should not suddenly take over every similar sound. A correction you repeat three times probably should.
Done carefully, this can work without an interruption after every sentence. Voxor can collect a candidate, wait for another example, and then prefer that spelling the next time the sound appears.
Sometimes you were only rewriting
There is an obvious trap here. People change text for reasons that have nothing to do with recognition.
Suppose the transcript says:
Send it to Chris tomorrow.
You replace it with:
Ask Chris whether Friday works.
That does not teach us a new pronunciation. You changed your mind. If a dictionary treated the edit as training data, it would learn nonsense.
Compare that with changing “Send it to vector tomorrow” to “Send it to Voxor tomorrow.” The sentence stayed intact, the replacement was local, and the two words sound alike. That is much better evidence.
We look for those practical differences: how much text moved, whether the replacement sounds like the original, whether it has happened before, and whether the surrounding sentence survived. None of those clues is perfect, so one isolated edit should remain a suggestion rather than becoming a rule.
That restraint matters. A dictionary that learns too slowly is mildly annoying. One that learns the wrong thing feels haunted.
The same sound can mean two different things
Simple search and replace works until a name overlaps with an ordinary word. If someone on your team is called Summer, you do not want every mention of the season capitalized. If Mercury is a project name, it may still be a planet in another document.
The surrounding words help, as does the app under the cursor. A term used in a code editor may deserve different capitalization from the same sound in an email. Recent project vocabulary can also break a tie without forcing a global replacement.
A learned entry works best as a preference. When the audio is ambiguous and the context fits, Voxor can choose the familiar spelling. It should never behave like an unconditional macro running over finished text.
You should be able to inspect what Voxor learned
Automatic learning becomes frustrating the moment it is invisible. When a surprising spelling appears, “the model thought so” is not an answer.
The dictionary should show the terms it has picked up, where each one came from, and whether it has been used recently. You should be able to correct an entry, remove it, or turn learning off. If you repeatedly undo a learned spelling, Voxor should take the hint and back away rather than defending an old guess.
This is also why the number of entries is a poor measure of success. A large dictionary may simply contain more bad guesses. What matters is whether you stop fixing the same words, and whether the dictionary avoids touching everything else.
Shared terms need a higher bar
For a team, the payoff is larger. Add the company name, a product, or an internal acronym once and everyone can get the same spelling. The cost of a bad entry is larger too.
Personal corrections should stay personal until there is a good reason to share them. A team term can be approved directly, or earn trust after several people use it independently. Scope matters as well. The legal team’s vocabulary does not need to influence every engineer’s transcript.
The useful shared dictionary is a short, current list of words the group truly has in common. It is not a warehouse for every term anyone has ever typed.
The result we want is simple to describe: correct a troublesome name once or twice, then forget the dictionary exists. Voxor keeps explicit custom terms, can learn from the corrections and documents already involved in your work, and leaves those choices visible when you want to inspect them.
Related reading: Voxor for writers, dictation software, and Local Mode.