Why does my iPhone voice memo transcript get names wrong?

It is a specific and repeatable annoyance: a transcript reads perfectly clean for three paragraphs, gets every ordinary sentence right, and then turns a colleague’s name into three different wrong words across three different mentions in the same recording. It rarely feels random once you notice the pattern, and it is not random — it is what a general-purpose speech model does when it meets a word it was never built to expect.

A recognizer does not know names exist

Speech recognition on iPhone works by matching the sound of what you said against a language model, then picking the most probable sequence of words that fits that sound — probable according to how the language is normally used, not according to who is actually in the room with you. That model has never met the specific people, products, or places in your life. It has only seen enormous amounts of ordinary language, and it scores every possible word against how often that word tends to show up in that kind of language.

A common first name usually appears often enough in that training to have decent odds. A less common name, a surname, a brand, or a made-up product name is rarer or entirely absent from what the model has seen, so its odds are low from the start — and the model still has to output something. It reaches for whatever common word or more familiar name sounds close enough to fill the gap, because "the closest thing I actually know" beats "nothing" every time recognition has to commit to an answer.

Why names specifically lose that contest

They are rare by definition

A word being uncommon is not an edge case for a language model, it is the normal condition for a huge share of real names. The whole reason a name identifies one particular person or thing is that most names are not shared by everyone — which is exactly the property that makes them statistically unlikely candidates for a system built on how likely a word is to appear.

They get little help from the words around them

Recognition leans heavily on context: an ambiguous sound in the middle of an ordinary sentence gets resolved because only certain words make grammatical and semantic sense in that slot. "This is [name] calling about the invoice" tells the model that the slot holds some kind of name, but grammar alone does not narrow it down to which one — every name is grammatically interchangeable there, so the model falls back on which name it has seen most, not which one you actually said.

It is not only people

The identical mechanism mangles brand names, medication names, street names, model numbers spoken as words, and technical terms specific to one field. Anything that is unusual enough to matter is, by that same trait, unusual enough for a general-purpose model to guess wrong about.

This is not a sign anything failed

A wrong name sitting inside an otherwise accurate transcript means recognition worked and made a specific, explainable kind of mistake — it is a different situation from a transcript that comes back blank, which means recognition failed to produce usable text at all. The audio itself is untouched either way; a mangled name in the transcript does not mean the name was recorded wrong, only that converting that particular sound to text landed on the wrong word.

It is also not something one recognition pass fixes and another does not. On-device recognition and a server-based fallback are both general-purpose language models trained the same way, on the same kind of ordinary language, rather than a lookup against a personal directory of names either one somehow has and the other lacks. There is no real basis for expecting a name to come through cleaner because it went through one pass instead of the other.

What narrows the gap, a little

• Say the name at a normal, unhurried pace rather than rushing past it — recognition works from the acoustic signal alone, and a name spoken quickly or trailed off at a lower volume gives it less to work with than the sentence around it.
• Use the same form of the name consistently within a recording. A model that has already committed to a wrong guess for one pronunciation is not helped by then switching to a nickname or a different pronunciation for the same person.
• Expect the same name to come out wrong the same way each time, not randomly. That consistency is actually useful: a name that is always turned into the same incorrect word is a quick, predictable find-and-replace once you notice the pattern, rather than something to hunt for differently every time.

Fixing it after the fact is the real fix

There is no setting that teaches a phone’s speech recognizer a specific name ahead of time, and there is no version of this that gets solved permanently before you record. The practical path is the same one that handles any other transcript mistake: read the transcript once, correct the name where it is wrong, and move on. That correction happens on the recording’s own entry, which is why it is worth doing before you export or share the transcript rather than after — a TXT or JSON file created from the text is a snapshot at that moment, and a correction made afterward does not reach a copy that already went out.

How Voice Studio handles this

Voice Studio transcribes through Apple’s Speech framework: on-device recognition first, and only a retry with on-device recognition switched off, sending the audio to Apple’s server, if that first pass comes back empty or errors. Neither pass has any special knowledge of the names in your life, so a mangled name is an ordinary outcome of how general-purpose speech recognition works, not something specific to one recording or one pass. The transcript on any recording is editable at any time, at no cost — correcting a name is the same free action as correcting anything else in it — and exporting as TXT or JSON afterward carries the corrected text, since export always reflects whatever the transcript reads at that moment.

Common questions

Is there a way to teach the app a name so it stops getting it wrong?

No. There is no custom dictionary or name list a recognizer checks first — every pass is a general-purpose language model with no knowledge of the specific people, places or products in your life. Correcting the name in the transcript after the fact is the actual fix.

Does the server-based fallback recognize names better than on-device recognition?

Not reliably. Both are general-purpose language models trained the same way, not a lookup against a personal directory, so there is no real basis for expecting one to handle a specific name better than the other.

Will transcribing the same recording again fix a wrong name?

Usually not. The audio has not changed and neither has the model reading it, so the same name tends to come out the same wrong way again unless something else about the conditions changed.

If I correct a name in one transcript, does it get recognized correctly in future recordings?

No. A correction applies to that recording’s own transcript entry. It is not a dictionary entry the recognizer consults for anything recorded afterward.