A voice note looks like one action and is actually four: the microphone records, the file is stored, a speech recognition model converts audio to text, and something indexes that text so it can be found later. Knowing where the seams are explains most of what people find surprising—why transcripts fail on names, why accuracy collapses in a café, and why an app that transcribes is not necessarily an app where you can find anything afterwards.
Step one: what recording actually captures
Your phone samples the air pressure at the microphone thousands of times per second and stores the numbers. Speech recognition typically works on audio downsampled to 16,000 samples a second in a single channel, because almost all the information that distinguishes speech sounds sits below 8 kHz. Music recording uses 44,100 and higher; speech does not need it.
This is why file sizes for voice notes are small and why microphone placement matters more than microphone quality. A phone in a pocket records mostly fabric, and no downstream model recovers what the microphone never received.
Step two: audio to text
Modern speech recognition maps audio directly to text using a neural model trained on large collections of transcribed speech, replacing the older pipeline of separate acoustic, pronunciation, and language models. The practical consequence is that the model predicts a plausible sentence rather than decoding sounds one by one.
That is why errors look the way they do. Recognizers rarely produce gibberish; they produce fluent, wrong words—a colleague's surname becomes a common noun that sounds similar, a product name becomes an English phrase. The output is always a complete-looking sentence, which makes mistakes harder to notice than static or silence would be.
How accuracy is measured, and what to expect
The standard metric is word error rate: insertions, deletions, and substitutions divided by the number of words actually spoken. A 5% WER means one word in twenty is wrong—usually acceptable for search, occasionally fatal for a number or a name.
Broad expectations, which vary widely by system and setup:
- Clear speech, quiet room, common accent: low single-digit error rates on general vocabulary.
- Background noise, several speakers, or distance from the microphone: error rates rise sharply, often several times over.
- Proper nouns, technical terms, and mixed-language sentences: the weakest area by a wide margin, because rare words are poorly represented in training data.
- Numbers, dates, and spellings: formatted by convention, so ‘fifteen hundred’ may return as 1500 or as words depending on the system.
Step three: making a transcript findable
Transcription and retrieval are separate problems, and plenty of apps solve only the first. If the transcript is stored per recording and searched by exact keyword, you can only find it by remembering a word you actually said—which, months later, is precisely what you have forgotten.
Retrieval by meaning changes the requirement: a memo where you said ‘the supplier keeps slipping deadlines’ is reachable from ‘vendor reliability problem’. That is the difference between a recorder and a searchable archive, and it is invisible in feature lists because both will claim transcription.
Why speaking beats typing for capture
The gap is mechanical. Speech runs at roughly 130 to 150 words a minute in conversation; phone typing runs at roughly 35 to 40 for most people. A ninety-second walk to the car can hold a complete thought that would not survive being typed with one thumb.
The trade is precision. Spoken notes ramble, restart, and leave the subject implicit—‘he said it would not work’ is clear to you for about a day. The cheap fix is a single closing sentence naming the person, the topic, and the reason you are recording at all.
What voice notes will not do
They will not correct themselves. A wrong name in a transcript stays wrong unless something flags the uncertainty or you catch it—and the fluency of the output actively discourages checking. Keeping the original audio alongside the text is the only real remedy: the recording is the record, and the transcript is an interpretation of it.
They also do not solve the case where you never speak the context out loud. And if a language you speak is not well represented in a model's training data, expect materially worse results in that language than in English, regardless of what the marketing says.
Treat the transcript as a searchable index and the audio as the source of truth. Test any voice notes app with one messy recording—background noise, two names, a number—then search for it a week later using words you did not say.