Voice Memo to Text

Q: Does it work on a phone?

The page fits a phone, but the work is heavy for one. Use whisper-tiny and short memos, or move the file to a computer for anything long.

What this page does

A voice memo is a recording, not a link, so this is a voice memo to text tool that starts where the memo starts: with the file the phone wrote. Drop it in and the spoken words come back as text, one timestamped line at a time, with SRT and VTT available from the same run. The engine is OpenAI's Whisper compiled to WebAssembly, running inside this tab, and the audio is decoded here too.

That last part is not decoration. Because the recording is read on this device rather than posted to a queue, the numbers below could be measured on the page itself, one file at a time, and they are the reason this page can tell you what a memo has to clear before it transcribes at all.

Getting the memo off the phone

The step before transcription is the one people get stuck on, and it is different on each phone. On an iPhone, open Voice Memos, tap the recording, tap the share button, then choose Save to Files or AirDrop it to a computer; if the memos sync through iCloud, the Voice Memos app on a Mac will also let you drag one out to the desktop. On Android, the recorder app's share or export action usually offers the recording as an audio file, and what it hands you is normally M4A, MP3 or WAV.

Either way you end up with a file, and the file is what this page wants. There is no server behind it, so it cannot reach into iCloud, a chat, or a cloud drive to fetch a memo for you — the memo has to land on the device first. That is the cost of the next paragraph being true: nothing is uploaded, so there is no copy of a private recording sitting anywhere afterwards.

The level a memo has to reach

Every page that answers this query talks about accuracy, and none of them says how loud a memo has to be. That is the gap worth closing, because a voice memo is the one recording most likely to be made badly: the phone on a table, in a pocket, in a car with the window down. So it was measured, on this page, with one recording and six copies of it.

The method. One 30-second stretch of speech was copied six times, and the only thing changed between copies was amplitude. Each copy went through the box above on whisper-tiny.en, in a fresh browser context, on 2026-10-18. The level of each file was measured with ffmpeg rather than read off the page, so the number in the left column is the file's own peak.

0.871 of full scale (−1.2 dBFS): 8 cues, 88 words, Done.
0.0871 (−21.2 dBFS): 8 cues, 87 words.
0.0275 (−31.2 dBFS): 8 cues, 87 words.
0.0219 (−33.2 dBFS): 4 cues, 44 words — the last timestamp is 00:00:11.502, and the page said Done.
0.0174 (−35.2 dBFS): refused, one sentence, no panel.
0.0087 (−41.2 dBFS): refused, the same sentence.

Where the line is. The page treats a file as speech only when its loudest sample clears 0.02 of full scale, which is about 34 dB below the top. Under that it does not ask the model anything; it answers with one sentence — No speech found in this file. It may be silent, music, or too quiet to make out. — and leaves the transcript panel closed. A memo built to look like a phone in a bag (dull, low-passed at 3 kHz, with the room noise it was recorded with) measured 0.0102 and got that sentence.

What happens just above it is the part nobody publishes. At 0.0219 the file cleared the line, so the page started, and it stopped again at 11.5 seconds of a 30-second memo. Four cues came back, the remaining nineteen seconds produced nothing, and the page still reported Done. The reason is in the shape of the check rather than in the model: the level of each stretch is tested before that stretch is sent, and a stretch under the line is skipped instead of guessed at. So a memo that starts close to the phone and then moves away is transcribed as far as the move, and the transcript simply ends there.

That gives you a diagnostic you can run without a meter: compare the last timestamp in the transcript with the length of the memo. If they match, the memo was heard to the end. If the transcript stops early, the rest was too quiet rather than misheard, and the fix is the recording, not a second attempt at the same file. A quick test before uploading: play the memo on the phone at the volume you normally use. If you have to hold the phone to your ear to follow it, it is under the line.

What is inside the .m4a

A file built to the shape an iPhone writes — mono, 44.1 kHz, AAC at 64 kbit/s in an MP4 container — was inspected with ffprobe before it was transcribed. It holds one stream, and that stream is audio. There is no subtitle track, no text track, no data stream: the count of non-audio streams is zero. Whatever words the phone displayed while you were recording, they are not travelling inside the file you export.

Two consequences follow, and both are useful. The first is that every route from a memo to text is a fresh listen — nothing can be extracted, only transcribed again — so the question is never "which tool reads the transcript" but "which tool hears the audio well". The second is that the container was not the obstacle in that comparison. The same 30 seconds delivered as that m4a and as a WAV came back with the same eight cues and the same 88 words, so for those two files the level was the variable, not the format. The codec inside the container is a different matter: an .m4a carrying Apple's lossless ALAC was refused by the browser's decoder before any transcription started, while the same recording as AAC came back normally — measured on this site on 2026-10-19, in Chrome. If a memo will not open, converting it is the fix; renaming it is not.

A memo is short until it is not

Most memos are under a minute, and a minute comes back as a handful of cues. The ceiling that produces them is fifteen seconds: a stretch of unbroken speech is cut at the fifteen-second mark, which is why a 30-second memo arrives as eight short cues rather than two paragraphs, and why a one-hour recording is roughly 240 of them. Cue boundaries otherwise follow the pauses, so a sentence usually ends where a cue ends.

The real limit on a long memo is memory, not time, because the whole recording is decoded before transcription starts. A three-hour recording on a laptop with little free memory is a genuine risk and on a phone it is a bad idea. Splitting the file is the usual answer and the cut has to land in a pause — measured on this site, a recording cut at a real pause kept every word across both halves, while the same recording cut at an arbitrary second lost the two words under the cut. Each piece also restarts at 00:00:00.000, so the second piece's timestamps need the first piece's length added back if you are lining them up against the memo.

Supported input formats

M4A, MP3, WAV, AAC, OGG and FLAC, plus the common video containers when the memo came with pictures — a screen recording, or a video someone sent. That list describes what browsers already decode rather than a filter applied here. A container the browser cannot read fails with a message in words instead of an empty panel.

Accuracy

No percentage is printed on this page. Word error rate moves with the microphone, the room, how far the phone was from the mouth, and how fast the person spoke, and one headline number would describe none of those. What can be said precisely: the models are the official Whisper checkpoints converted to ONNX by the Xenova project — tiny, base and small — and a larger one generally writes more of the line, at the cost of memory and time.

For a memo, three things are worth checking by eye before you trust the rest: names, numbers, and the last timestamp. Names and numbers are where a speech model substitutes something plausible. The last timestamp is where a quiet memo gives itself away.

FAQ

Is the memo uploaded to a server?

No. The audio is decoded in this tab and the Whisper model runs here as well, compiled to WebAssembly. There is no backend that receives the file. Four things do leave the browser: the one-time download of the model weights, a cookieless page-view count sent to Cloudflare Web Analytics, the Google Analytics script this page loads, and the visit reports that script sends to Google. The privacy policy names each one.

Why did my transcript stop partway through?

Because part of the memo fell under the level line. The page checks the loudest sample in each stretch before it sends that stretch to the model, and a stretch under the line is skipped rather than guessed at. Measured on 2026-10-18: one 30-second memo, its loudest moment at 0.0219 of full scale, came back as four cues ending at 00:00:11.502 — the rest was dropped, and the page still said Done. Compare the last timestamp in the transcript with the length of the memo: if they do not match, the missing part was too quiet rather than mistranscribed.

How loud does a voice memo have to be?

The loudest moment in the file has to clear 0.02 of full scale, which is about 34 dB below the top. That is the line the page uses to decide whether there is speech to transcribe at all. Measured on 2026-10-18 with one 30-second recording and six copies of it, the result changed at exactly that point: 0.0275 of full scale returned the whole memo, 0.0174 returned one sentence saying no speech was found. A memo recorded with the phone in a pocket sat at 0.0102 and was refused. If you have to hold the phone to your ear to follow a memo, it is under the line.

Do I need to convert the .m4a first?

No. An iPhone voice memo is already an MP4 container holding AAC audio, and this page reads it as it is. Measured on 2026-10-18: the same 30 seconds of speech as a mono 44.1 kHz AAC file at 64 kbit/s and as a WAV returned the same eight cues and 88 words. Converting to MP3 only loses a little quality.

Does it work with Android recorders?

Yes. An Android recorder usually hands you M4A, MP3 or WAV, and each of those is read here as it is. What decides the result is the same thing it is on an iPhone: the level the recording reached.

Can it tell who is speaking?

No. There is no speaker diarization here: the words arrive as one stream, and at a change of speaker the text simply continues. Cue boundaries follow the pauses, so a new voice usually starts a new cue, but no name is attached to it.

Can it transcribe a voice note someone sent me?

If you can get the file onto this device, yes — it is transcribed like any other recording. What this page cannot do is reach into a chat or a cloud folder and fetch it: there is no server here, so the file has to arrive first.

How long a memo can it take?

No length is set by the page; the limit is the memory of the device, because the whole recording is decoded before transcription starts. A one-hour memo is roughly 240 cues at the fifteen-second ceiling. Splitting it is the usual answer, and the cut has to land in a pause: measured on this site, a recording cut at a real pause kept every word across both halves, while the same recording cut at an arbitrary second lost the two words under the cut.

Does it work on a phone?

The page fits a phone, but the work is heavy for one. Use whisper-tiny and short memos, or move the file to a computer for anything long.

What comes out, and can I get subtitles?

TXT with a timestamp at the start of every line, and SRT or VTT from the same run. The subtitle formats were checked against outside readers rather than assumed: the exported files were read back by three parser libraries and ffmpeg, cue for cue.

Privacy

The memo stays on this device. There is no account, no upload step and nothing to delete afterwards — closing the tab ends it. The visit itself is counted twice, by Cloudflare Web Analytics, which sets no cookies, and by Google Analytics, which keeps _ga and _ga_KJ155PR9B7 in this browser for two years. The privacy policy lists what each records.

This matters more for a memo than for most audio. A voice memo is the recording people make when they are thinking out loud, rehearsing something difficult, or remembering a conversation they did not want to write down. It is the file least suited to being posted to a queue and least likely to be deleted from one.