The dictation packages now claim accuracy in the high nineties, and that figure is quoted per word, which is the detail that makes it much less impressive than it sounds.

Ninety-five percent per word is one error every twenty words. An average sentence runs to about twenty words. So a system advertised as extremely accurate is producing roughly one mistake per sentence, and you must read every sentence to find it, which means proofreading everything you dictated. At that point the labour has moved rather than disappeared.

Ninety-nine percent is one error in five sentences, which is genuinely usable. The gap between those two figures is enormous and the marketing presents them as neighbours.

There is a second and larger problem that better recognition does not touch, which is that transcription is not comprehension.

Suppose the machine captures every word perfectly. It now holds a string of text and still has no idea what you meant by it. Ask a phone to ring somebody and you have to say the name in the form it expects, in the order it expects, with a pause where it expects one. Get any of that wrong and you have not made an error of pronunciation, you have made an error of protocol, and the device cannot tell you which.

So what looks like a speech problem is mostly a language problem wearing a speech problem’s clothes, and it is much further from being solved.

Then there is the thing nobody in the field wants to discuss, which is that talking to a machine in front of other people is embarrassing.

Not slightly awkward. Genuinely uncomfortable, in an office, on a train, in a queue. People will silently type an entire paragraph on a numeric keypad rather than say six words aloud where colleagues can hear, and no improvement in accuracy addresses this, because the obstacle is social rather than technical.

Which points at where this actually wins. Driving, where your hands are occupied, nobody is listening, and the vocabulary is small enough that understanding is tractable. Perhaps a kitchen. Perhaps for people who cannot use a keyboard, where the comparison is not against typing but against nothing.

Everywhere else the keyboard keeps winning, and it will keep being described as a failure of the technology rather than a preference about being overheard.