Tools
Transcription
Recordings made with the Voice Memos app on an iPhone, or with a handheld IC recorder, are saved with the ".m4a" extension. Some transcription services refuse to accept that extension and tell you to convert the file to MP3 first, but whether a conversion is needed depends entirely on the tool you picked, not on m4a being an awkward format.
This article walks through transcribing an m4a file without converting it, starting from moving the recording onto a PC and going all the way to separating speakers and exporting a subtitle file. No account and no app installation are required, and everything runs in the browser.
Here is what you need and what you do not.
- What you need — a PC, Chrome or Edge, and a ".m4a" recording
- What you do not need — an account, an app install, a file upload, or a format conversion
m4a can be transcribed as it is
"m4a" is the container used when audio is compressed with AAC, and it is the default save format for the iPhone Voice Memos app, for several Android recording apps, and for most IC recorders. As audio data it behaves the same way MP3 does, so as long as the tool supports it, no conversion is necessary.
You are asked to convert when the library that the service runs internally cannot read m4a. The tool I built, "Transcriber", runs ffmpeg inside the browser, so it pulls the audio out of an m4a or an MP4 the moment the file is loaded.
There are several ways to transcribe a recording, and the choice changes both how large a file you can handle and where your audio ends up, so it is worth comparing them first.
Method | Conversion | Length limit | Speaker separation | Where the audio goes |
|---|---|---|---|---|
Built into the iPhone | Not needed | None | Not available | Stays on the device |
Cloud services | Depends on the service | Capped on free plans | Usually a paid feature | The provider's servers |
Processed in the browser | Not needed | None | Available | Never sent anywhere |
Desktop software | Not needed | None | Configurable | Never sent anywhere |
If you are on an iPhone and reading the text on that same device is enough, the built-in feature is the quickest route. Apple brought Apple Intelligence to Japanese in iOS 18.4, so on a supported model the Voice Memos app can turn a recording into text on its own.
The built-in feature cannot split the text by who was speaking, however, and it has no way to export a subtitle file, so for a meeting or an interview with several participants the later work goes much more smoothly if you process the recording on a PC.
What to prepare
- A PC. Windows, macOS and Linux all work
- The latest Chrome or Edge. Safari and Firefox work too, but they are slower for the reason described below
- A ".m4a" recording. Files up to 2 GB can be loaded
Step 1: Move the recording to a PC
First copy the file from the recording device onto the PC. The steps differ by device, so follow the one that matches your setup.
- On an iPhone, press and hold the recording in Voice Memos and choose AirDrop or "Save to Files" from the share menu. AirDrop is fastest on a Mac; on Windows, use iCloud Drive or a USB connection.
- On Android, share the file to Google Drive from the recording app, or connect over USB and open the recordings folder directly.
- With an IC recorder, connect it over USB and it appears as external storage, so you can copy the file straight across.
Step 2: Transcribe it in the browser
Once the file is on the PC, open a transcription tool in the browser. I am using "Transcriber", which I built myself, but the shape of the procedure is the same for any tool that reads m4a directly.
Transcribe audio and video files entirely in your browser. No account and no upload — your audio never leaves your device. Free with no time or usage limits. Speaker diarization and SRT subtitle export powered by Whisper AI, supporting MP3, WAV, M4A, MP4 and more.
- Open "Transcriber" in the browser.
- Set "Model", "Language" and "Speaker diarization" before doing anything else.
- Drop the m4a file onto the area marked "Drop an audio or video file, or click to select".
- Processing starts as soon as the file is loaded. Wait for it to finish.
Choosing a model
Six models are available. Larger ones transcribe more accurately, but they also take longer and use more memory.
Below are the numbers I measured earlier on a 4 minute 08 second video (51.9 MB). The PC had an RTX 2060 SUPER (8 GB), the browser was Chrome, and WebGPU was enabled.
Model | Processing time | Speed relative to real time |
|---|---|---|
Small | 0 min 35 s | about 7.1x |
Medium | 1 min 22 s | about 3.0x |
Large v3 Turbo | 0 min 37 s | about 6.7x |
Large v3 | 0 min 52 s | about 4.8x |
People assume that a bigger model is always slower, but these numbers do not bear that out, and the slowest of the four was the mid-sized Medium. Large v3 Turbo finishes almost as fast as Small because the part that analyses the audio is the same size as in Large v3, and only the part that assembles the text was made smaller.
I use "Large v3 Turbo" myself. How reliably it picks up Japanese proper nouns and technical terms is clearly different from Small, and since it takes roughly the same amount of time, I see no reason to choose anything else.
Step 3: Separate the speakers and export
When the recording has more than one person in it, turn "Speaker diarization" on before starting. The exported text is then split line by line according to who said what.
If you already know how many people took part, set it in "Number of speakers (0 = auto)". Automatic detection works, but telling it the number improves how accurately lines are assigned.
This feature is a port of the "pyannote.audio" speaker diarization pipeline to the browser; it slices the audio into ten-second chunks, extracts a fingerprint for each speaker and groups the segments that belong to the same person. Very short interjections of about a second are sometimes missed, so those are worth fixing on screen after the export.
Choosing between text and subtitles
Two export formats are available, "TXT" and "SRT (subtitles)", and which one to pick depends on what happens next.
- Use TXT for meeting notes or a draft of an article
- Use SRT for adding subtitles to a video
When it does not work
When processing stops partway through, the failure reports that reach me point almost entirely at one cause, which is running out of memory.
- Smartphones stop partway through more often, and the file size limit there is 500 MB
- Where WebGPU is unavailable, such as Safari and Firefox, the work falls back to the CPU and becomes far slower
- For recordings over an hour, start with a small model and move up only if the accuracy is not good enough
If the warning "WebGPU is not available in this environment, so this model will be very slow" appears, reopen the page in Chrome or Edge, or pick a smaller model if that is not an option.
Frequently asked questions
Is it really free? Are there limits on length or on how many times I can use it?
It is free. There is no limit on length and no limit on the number of runs, so no charge appears however many hours you load or however often you do it. The transcription runs on the visitor's own PC, which means no server costs land on my side.
Does accuracy improve if I convert the m4a to MP3 first?
It does not. Converting decodes the audio and then compresses it again, so the quality drops rather than improves. Load the m4a as it is.
Is there any chance the recording is sent to a server?
There is not. The transcription runs entirely inside the browser, so the audio file never leaves the device. The only network traffic is the initial download of the AI model.
Do I need an account or a login?
You do not. No email address is asked for either, so you can start using it the moment the page opens.
Can I do all of this on an iPhone or an Android phone?
You can, but I would not recommend it. Phones have far less memory available, so loading a long recording often stops partway through. Recording on the phone and transcribing on a PC is the dependable combination.
How long a recording can it handle?
There is no limit on length as long as the file stays under 2 GB. Longer recordings take proportionally longer to process, so the tab has to stay open until it finishes.
Summary
An m4a file can be transcribed as it is. Being told to convert it to MP3 first only means that the service in question cannot read the format, not that there is anything wrong with the audio itself.
- Move the recording to a PC
- Set the model, the language and speaker diarization first
- Load the file (processing starts the moment it is loaded)
- Export as TXT or SRT
Related articles and tools
A step-by-step guide to turning a recorded X (formerly Twitter) Space into text for free. Covers downloading the recording, separating speakers and exporting subtitle files, all in the browser without an account or an app install.
No install, no upload. Transcribe a video, fix the SRT, and burn the subtitles into the picture — all inside your browser, free and without an account.
Convert audio to MP3, WAV, M4A, AAC, OGG, OPUS, FLAC or AIFF, or extract audio from MP4 and MOV videos. Free, in your browser, nothing uploaded, no length limit.
media.tools.ryusei.io
Programing
Python
Hosting