Code dictation practice drills

Why fifteen phrases is the whole learning curve

VoiceGem recognizes 90 spoken phrases, and almost nobody uses more than about fifteen of them. Parentheses, brackets, braces, dot, comma, semicolon, equals and a casing command cover the overwhelming majority of what appears in a dictated line, and the rest are there for the day you need an ampersand.

The reason drills work better than reading the table is that the phrases have to become automatic. A phrase you have to recall is slower than typing the character, which is the whole argument against code dictation; a phrase you say without thinking is not. Twelve short drills is roughly where the recall stops being conscious.

The drills are marked by running your attempt through the real formatter rather than comparing it to a stored answer string. That means a drill cannot pass here and fail in your editor, and it means any valid alias works — "open paren", "open parenthesis" and "left paren" all pass the parenthesis drill, because all three produce the same character.

The order the drills are in, and why

Single symbols come first, because spacing is the thing that surprises people. A period attaches to the tokens on both sides, so "user dot name" gives user.name with no gaps. A semicolon attaches left and breathes right. Getting a feel for those two flags early makes everything after them predictable.

Casing commands come next, because they behave differently from symbols: a casing command claims the run of ordinary words that follows it and stops at the first symbol or command. That is the one rule in VoiceGem that is not a straight substitution, and it is the rule people trip over when a command swallows more of the line than they expected.

Compound lines come last. By that point the individual phrases are automatic, and the only new skill is saying a whole line without pausing in the middle — which matters, because a pause in the wrong place gives the transcription model a sentence boundary you did not intend.

What practice does not fix

Per symbol, saying a phrase is slower than typing the character, and no amount of practice reverses that. Dictation wins on whole utterances — a function signature you would type in forty seconds takes eight to say — and loses on single-character edits. The right habit is dictating the first draft and typing the fixes, and drills that pretend otherwise would be training the wrong reflex.

Practice also does not help with vocabulary. A library name a general speech model has barely seen will come back wrong however clearly you say it, because the model is choosing the more probable transcription rather than mishearing you. That is a replacement rule's job, and the per-project vocabulary builder on this site is where to collect them.

And no drill makes an open-plan office quieter. VoiceGem needs you to talk, and if the desk you work at makes that socially expensive, the honest answer is that these drills will not change it.

How to practice against your own code

After the twelve, the useful exercise is your own repository rather than more drills. Open a file you wrote last week, pick three lines, and work out what you would say for each — the speech-to-code converter on this site will tell you whether you were right, on lines that look like the ones you actually write.

Lines that resist are worth noticing rather than forcing. Dense generic signatures, regular expressions and anything with a run of adjacent symbols are genuinely slower to dictate than to type — VoiceGem does not change that, and knowing which of your lines fall into that category is more useful than getting faster at saying them.