Voice input stopped being a feature and became an interface

If you tried dictation five years ago and gave up, the reason you gave up was probably not that the words came out wrong. It was that using it was a detour.

The thing that changed is not accuracy

Speech recognition crossed the threshold of being good enough for ordinary office English some time ago. Whatever the remaining gap is between the best system and a careful human transcriber, it is not the gap that decides whether you use dictation on a Tuesday afternoon. Most people who have abandoned voice input did not abandon it because it misheard them. They abandoned it because of everything that had to happen around the speaking.

The older shape of the tool went like this. You opened a dictation application. You spoke into it. You got back a block of raw text, complete with every false start, every “um”, every time you changed your mind halfway through the sentence. Then you selected that block, copied it, switched to the application you actually wanted to write in, pasted it, and began the real work of turning a transcript into prose. The recognition step was the fast part. Everything on either side of it was slow, and the slowness was structural rather than technical.

Two things have shifted. The first is that dictation moved out of its own window and became a system level input method, which means the text arrives where your cursor already is. The second is that a language model now sits between the audio and the page, cleaning up the speech before it lands. Neither of those is a recognition improvement. Both of them are changes to where the tool sits in the sequence, which is exactly the kind of change that decides whether software survives a real week.

Destinations and input methods

It is worth being precise about this distinction, because it explains a lot of software that otherwise looks similar on a feature list.

An input method has a much harder job and a much larger payoff. The job is harder because it has to work everywhere without knowing anything about where it is: in a chat box, in a code editor, in the subject line of an email, in a form field on a website nobody has heard of. It cannot ask for an integration. It has to behave like hardware.

The payoff is larger for the same reason. A destination tool has to be remembered. Every time you want it, you have to decide to go and get it, and that decision is where adoption dies. An input method does not need to be remembered in the same way, because it lives at the point where you were already about to type something.

The second pass is the actual product

Here is the part that is genuinely new, and it is easy to miss because vendors tend to sell it as an accuracy improvement when it is not one.

When you speak an unrehearsed sentence, you do not produce prose. You produce something closer to a first draft delivered out loud: you start, you stop, you back up, you insert a clarification in the middle, you say “sorry, I mean Monday”. A transcriber that is perfectly accurate will faithfully write all of that down, and the result is a mess that takes real effort to clean. Perfect accuracy is, in this specific sense, the wrong target.

What the current generation of tools does is run a second model over the transcript that strips fillers, resolves self corrections, adds punctuation and capitalisation, and in some cases adjusts register depending on which application the text is going into. That second pass is what removes the editing detour, and removing the editing detour is what makes speaking competitive with typing for real work rather than just faster in a words-per-minute sense.

Our judgement: this is the single most important thing to understand about the current crop of voice tools, and it is systematically underplayed in marketing. The cleanup pass is a real advance and it genuinely changes the economics of dictation. It also creates a new category of error that is harder to catch than the old one, because the old errors looked like errors.

What an input method has to earn

If speech is going to sit alongside the keyboard rather than beside it in a separate window, there is a short list of things it has to get right. None of them are about recognition quality.

1. It has to be instant

   The gap between deciding to speak and being able to speak has to be close to zero. A keyboard has no startup cost. Anything with a noticeable one loses to the keyboard for short text, which is most text.

2. It has to be reachable without looking

   Which in practice means a modifier key you hold, not a button you find. The moment you have to locate something on screen, you have switched contexts, and switching contexts is the cost you were trying to avoid.

3. It has to fail quietly

   An input method that occasionally dumps a paragraph of garbage into a live chat is worse than no input method. The cost of a bad activation is paid in public.

4. It has to be predictable about where the text goes

   Always at the cursor, never somewhere clever. The value of behaving like hardware comes entirely from being boring about this.

5. It has to be honest about what it sends

   Cleanup that is aware of the surrounding application is useful, and it is also a much larger disclosure than “we process your audio”. This is where most of these products are least clear, and it is worth checking per tool rather than assuming.

The last one is not a privacy footnote appended for form's sake. Context aware cleanup means the tool is reading something about the window you are in so that it can match the register or the vocabulary. How much it reads, and whether that leaves your machine, varies enormously between products and is usually documented somewhere other than the page that sells you the product. We went through one vendor's version of this in detail in our Wispr Flow feature.

Where this leaves the keyboard

Not replaced, and probably never replaced. The keyboard is precise, silent, private in an open office, and it does not care whether you have a cold. Its real weakness is narrow: it is slow at converting a thought you have already finished having into characters.

That is a genuine bottleneck and it is worth removing, but it is a smaller share of most people's writing than the four-times-faster framing implies. A long reply you have been putting off is mostly transcription. A short precise message with three proper nouns in it is mostly not. We wrote a separate piece on where that line falls, because it is the practical question and it does not have a single answer.

The more interesting prediction is not that voice replaces typing. It is that input becomes plural in the same way that pointing did. Nobody argues about whether the trackpad replaced the mouse. You use whichever one is in front of you, and the software does not need to know which.

What would make it stick

Three things, none of which are model improvements, and all of which we will be watching for.