Your transcript will have mistakes in it. That's fine.

The most common reason people quit a voice-first tool isn't accuracy. It's that they treat the transcript like a document. They read it back, find "Sarah" rendered as "Sara," find a garbled half-sentence, and start fixing — and now talking a journal costs more than typing one did, which is the whole problem it was meant to solve. Same failure mode we wrote about in why journaling habits keep dying: the entry starts costing more than it returns.

It helps to know exactly where the errors come from, because then you can tell which ones are worth a second of attention.

Where speech recognition actually fails

Recognition isn't uniformly unreliable. It's very good in the middle and weak at specific edges:

• Proper nouns. Names of people, streets, bands, companies, medications. A model predicts likely word sequences, and your coworker's surname is not a likely word sequence. This is the single biggest error category in personal recordings.
• Homophones and near-homophones. There / their, to / two, "a loud" / "allowed." Context resolves most, but a fragment with no context can flip.
• Trailing words. People drop volume at the end of a sentence. The last two or three words of a thought are the most likely thing in the entry to come out wrong or missing entirely.
• Numbers and dates. "Fifteen" and "fifty" differ by one unstressed syllable. "The 4th" and "the 14th" are close. Spoken dates are ambiguous even to humans.
• Crosstalk and noise. A TV, a car, someone else talking. Recognition degrades fast when two voices overlap.
• Code-switching and accents. Switching languages mid-sentence, or a strong regional accent against a model trained mostly on other accents, both raise the error rate — and this is unevenly distributed, which is worth being honest about rather than pretending everyone gets the same experience.

Why a journal forgives nearly all of it

Now compare that list to what a journal entry is actually for.

An email has an audience who wasn't there. A meeting transcript is evidence. A published article is judged on its prose. A journal entry is a memory trigger for one person who was present — and that person is reading it precisely because they already lived it.

If Thursday's entry says "the thing with Sara went sideways" and her name is really Sarah, you lose nothing. If a sentence trails into nonsense, you almost always remember what you were saying, because you said it. The information in a journal is carried by the gist, the ordering, and the emotional register far more than by any individual word.

Which means the accuracy bar for this format is much lower than the bar for dictating a work email — and people apply the email bar by reflex.

The four places errors genuinely cost you

1. Dates and times. An entry that moves an appointment to the wrong day is worse than no entry. Say them unambiguously — "Thursday, the fourth of September" beats "the fourth."
2. Amounts. Money, dosages, measurements. Fifteen versus fifty is a real error with real consequences, and it's one of the likeliest confusions there is.
3. Names you'll search for later. If you'll ever type a name into search to find this entry, the spelling has to be close enough to match. Say it clearly the first time it appears, and it'll usually anchor the rest of the recording.
4. To-dos. The extracted task list is the part of the entry you'll act on without rereading the context, so it's the one place a garbled word can quietly cost you something. Glance at the tasks. Skip the prose. More on that flow in our voice-notes-to-tasks guide.

The honest trade-off with on-device

Running recognition and summarization on the phone rather than a server has a cost, and it's worth stating plainly: a phone-sized model has less capacity than the largest cloud models, so on the hard edges above — unusual names, heavy background noise, overlapping speakers — the cloud generally has an advantage.

That trade is deliberate. The failures you're buying are cosmetic and clustered in the places a journal doesn't care about. What you're buying with them is that the most unguarded thing you produce all week never leaves the device — the argument in why a journal is the worst thing to put in the cloud. A slightly worse transcript of something honest beats a perfect transcript of something you edited for an audience you imagined.

How to shoot for a good transcript without proofreading one

1. Say names and dates deliberately, once. Slightly slower, slightly clearer, at the top. Everything else can be as sloppy as you like.
2. Flag your to-dos as sentences. "I need to call the dentist tomorrow" extracts far more reliably than a trailing "…oh, dentist."
3. Finish your sentences with volume. The single cheapest accuracy win, because trailing-off is the biggest avoidable error source.
4. Record somewhere reasonably quiet, or accept the hit. Walking outside is fine. A bar is not, and no model fixes that.
5. Read the tasks. Don't read the prose. If you catch yourself editing a sentence for style, close the entry — that's the habit dying in real time.

The framing that makes this easy

A journal is not a manuscript. It's a set of handholds you leave for a future version of yourself who was, conveniently, present for everything in it. Handholds tolerate typos. Manuscripts don't.

Get the four things right that a future you can't reconstruct — when, how much, who, what next — and let every other word be approximately correct. That's a working journal, and it takes two minutes.