How to fix a bad transcript on iPhone

A transcript that is ninety-five percent right and wrong in the same five percent every time is not a broken feature. It is what speech recognition actually is: a model guessing at sound, doing very well on ordinary words and badly on the words it has never been trained to expect.

Why the words come out wrong in the first place

Two different things decide what you get back. One is which recognizer ran: iOS speech recognition tries the on-device model first, and only falls back to Apple’s server if that local pass comes back empty or errors. Either path can produce a wrong word — this is not a case of one being accurate and the other not, it is that both are guessing under uncertainty, just with models of different sizes.

The other is what was actually in the room. Distance from the speaker, background noise, and people talking over each other degrade the signal before either recognizer ever sees it. A model can only guess at what it received, and a quiet mumble a metre from the microphone gives it very little to work with.

The errors that repeat, and why

Names and proper nouns

A speech model is trained on general language, not on your contacts list or your family. It has never heard your street, your colleague’s surname, or your dog’s name, so it substitutes the nearest ordinary word it does know. This is the single most common correction anyone makes to a transcript.

Jargon and technical terms

Same mechanism, different source. A model trained on everyday speech has no reason to expect a chemistry term, a legal phrase, or the name of your team’s internal tool, and it reaches for something that sounds similar instead.

Numbers, spellings and anything said quickly

A phone number, a spelled-out email address, or a fast list of items gives the recognizer less time per word to work with, and errors cluster there even when the rest of the sentence comes out clean.

Overlapping or quiet speech

When two people talk at once, or someone trails off at the end of a sentence, the recognizer is working from a genuinely degraded signal. No amount of correcting afterwards fixes that at the source — it just tells you where to listen more carefully when you check.

Two different failures, and they need different fixes

It helps to tell these apart before you start correcting, because they call for different amounts of work.

• A wrong word in an otherwise complete sentence. The recognizer heard something and guessed wrong. You already know roughly where it is — the sentence around it still makes sense — so this is a quick, targeted fix.
• A blank or garbled stretch. The recognizer returned little or nothing for that section, usually because the audio itself was too quiet, too noisy, or in a language whose local model struggled with that clip. Correcting this means listening to the audio and filling in text that was never generated, which takes longer and is worth doing only for the parts that matter.

How to fix it without redoing the whole thing

• Read the transcript at normal reading speed once, all the way through. Most of it will be fine — do not slow down for sentences that already make sense.
• Where something looks off — a word that does not fit the sentence, a name spelled unusually, a gap — play just that section of audio back rather than the whole recording.
• Fix only what is actually wrong. A transcript does not need to be beautiful, it needs to be accurate at the handful of places where it currently is not.
• If a name or term appears more than once, expect it to be wrong every time it appears, not just the first. The recognizer has no memory of your earlier correction.

Do this soon, not eventually

A transcript gets copied, searched and forwarded long before anyone thinks to double-check it. If you search for a name later and it does not turn up, the most common reason is that the recording mis-heard it and nobody has fixed the text yet — a mangled word is invisible to a search for the correct one. Correcting it once, while you still remember what was actually said, is what makes it findable and quotable from then on.

What editing cannot do

Editing the text fixes wrong words, not missing structure. iOS speech recognition does not label speakers, so a transcript of two people talking is continuous text, and no amount of correcting adds the "A:" and "B:" back in — that has to be done by hand if you need it. And if a stretch of audio came back genuinely empty because it was too quiet or too noisy, editing means listening and typing what you hear yourself; there is no button that makes a bad recording transcribe better the second time. If the audio itself is the problem, the real fix is upstream of the transcript: closer microphone placement and a quieter room next time.

How this works in Voice Studio

The transcript is editable after the fact — fix a name, keep the rest. That is a free-app feature, not something held back for Pro; correcting a transcript is part of ordinary use, the same as recording or exporting it. Once it is fixed, it stays fixed everywhere the text is used: the library search that looks at titles, transcripts and place names will find the corrected word, and a TXT or JSON export carries the correction along with it, not the original guess.

Common questions

Why does my transcript get the same name wrong every time it appears?

The recognizer has no memory between occurrences — it is guessing fresh each time based on the sound, not recalling an earlier correction. Fixing the first instance does not fix the rest; each one needs the same correction.

Can I get iOS to transcribe a blank section again?

Editing a transcript on iPhone means listening and typing the missing words yourself. There is no separate retry step for a section that came back empty — if the audio itself was too quiet or noisy, that limitation applies again on a second attempt.

Does correcting a transcript change the original audio?

No. Editing only changes the text. The recorded audio is untouched, so you can always go back and re-check a correction against what was actually said.

Does fixing a transcript cost anything in Voice Studio?

No. Editing the transcript is part of the free app, the same as recording, markers, search and export. It is not one of the things Pro adds.