Transcribe a phone call recording, privately

Works with the files your phone makes

Call recorders on Android mostly save one of a handful of formats. Look at the end of the file name to see which one you have:

The .amr extension covers two different codecs, and plenty of apps do not say which one they use. That does not matter here. An AMR file starts with a short text marker naming its type, and the tool reads that marker rather than trusting the extension. For a .3gp file, the conversion command is on the MP3 to text page, and it runs on your own computer, so converting does not mean uploading either.

Your recorder app's settings usually name the format it uses, and many let you change it. If yours offers a wideband or HD option, it is worth choosing, for reasons covered below.

Getting the recording from your phone to your computer

You can transcribe on the phone, but a computer is much faster. On a desktop with a graphics card, our 17-minute test call took about two minutes for the speech and a little over two more to sort out who was talking. A phone takes several times longer and has to keep its screen on throughout. So the usual route is to move the file first. Four ways, from most to least private:

• USB cable. Plug the phone in, choose File transfer in the notification that appears on the phone, and open the phone's storage from your computer. Recorder apps keep their files in a folder of their own, usually named after the app. The file never touches the internet.
• Quick Share. Google's Quick Share app for Windows lets you send a file from an Android phone to a nearby PC directly, without going through a cloud account.
• Google Drive. The most convenient if you do this often. Share the recording to Drive from the recorder app, then pick it on the computer. With Google's Drive for desktop installed, your Drive shows up as an ordinary drive, usually G: on Windows, so you can choose the file straight from G:\My Drive in the tool's file picker without downloading it first. The trade-off: Google now holds a copy of the call. MeetingScribe still uploads nothing, but the recording is no longer only on your devices.
• Email it to yourself. Works anywhere, with the same trade-off as Drive, and long calls can exceed attachment limits.

Once it is on your computer the file still will not open in most desktop players, which matters if you want to keep or share the recording itself rather than the transcript. TidyAudio turns it into an MP3 in the browser, and can trim the dead air off the ends while it is there. It is a sister site of ours and works the same way, with nothing uploaded. One difference worth knowing: it reads ordinary narrowband .amr only, so if your recorder saved a wideband .awb file, transcribe it here and keep the original as it is.

Why call recordings are harder to transcribe

A call recording is one of the toughest inputs a speech model gets, for three reasons that stack on top of each other.

The other side is a second-hand copy

Recent versions of Android do not let a recorder app you install yourself capture the call audio directly. Those apps record through the microphone instead. Your own voice goes straight in; the other person only arrives as the sound leaking out of the earpiece next to your cheek. The result is the familiar lopsided recording: you loud and clear, them faint and thin. On speakerphone the gap mostly closes, because their voice then comes out of a loudspeaker the microphone can hear properly. A recorder built into the phone's own call app is a system app with access to the call audio itself, so it tends to produce a more even recording.

Phone calls are narrowband

A traditional call carries only the range of sound from roughly 300 to 3,400 hertz, and narrowband AMR stores audio at 8,000 samples a second, so nothing above 4,000 hertz survives. That is fine for a human, who fills in the gaps from context, but the hiss of an s, the friction of an f and the breath of a th all live largely above that line. The model then has to guess between "fifty" and "sixty", or "fees" and "seas", from what is left.

HD Voice calls, which most phones and networks now use when both ends support them, are wideband: roughly 50 to 7,000 hertz, stored by AMR-WB at 16,000 samples a second. That happens to be the exact rate the speech model reads, so a wideband recording gives it everything it can use. One catch: a wideband file does not upgrade a narrowband call. If the call itself was carried at phone quality, perhaps because the other person was on an older network, their side stays band-limited inside the wideband file, while your own voice, captured by the microphone, keeps the full range.

It has been compressed twice

The network compresses the other person's voice to carry it, and the recorder compresses it again to save space. AMR is built to keep speech understandable at very low bit rates, around 12 kilobits a second, and it gets there by modelling the voice rather than storing the sound wave. Each round loses a little of the fine detail that separates similar words.

What helps

• Record on speakerphone when the call matters. It is the single biggest improvement for the other person's side.
• Choose wideband if your recorder offers it, and keep VoLTE switched on in the phone's settings (some phones call it HD Voice or Advanced Calling).
• Tell the tool what the call is about. Names and trade terms are what phone audio garbles most. A sentence such as "Call with Priya Nair at Northgate Plumbing about the backflow test and the permit" in the What is this recording about? box lets the tool correct those words when they are misheard.

MeetingScribe also does two things on its own:

• It restores the consonant range on phone audio. Before transcribing, the tool measures whether the recording is phone quality. If it is, it lifts the frequencies where consonants live, and tells you it did. On a real 21-minute call this took the transcript from 2,502 words to 2,614 and fixed industry terms the untreated audio got wrong. Recordings that are not phone quality are left alone.
• It offers a volume fix for lopsided calls. The Also even out volume setting brings a faint voice up to the level of a loud one. It is off by default because on balanced recordings it made things worse, but a call where you can barely hear the other side is exactly what it is for.

After transcribing, Refine with your AI hands the transcript to the AI chat of your choice to correct words using the context of the whole conversation. That catches the phone-audio mistakes the speech model cannot, and you approve each change. The guide to fixing misheard words shows how it works.

Telling you and the other person apart

Turn on speaker labels and set How many people talking? to 2. Leaving it on Work it out for me is fine for meetings, but on a call you already know the answer, and stating it removes the guesswork: on our own 21-minute test call, the automatic count found the right two people when it ran on the graphics card and three when it ran on the processor. The transcript comes back with SPEAKER_00 and SPEAKER_01, which you can rename to yourself and the caller in one step.

Calls are full of short replies, such as "yeah", "right" and "okay, go on", and those are what speaker labelling most often hands to the wrong person. The Sharper speaker labels setting times each word separately, so those short replies stay with whoever said them, at the cost of a longer run. How labelling works is explained on the speaker identification page, and what to do when labels come out wrong is on fixing speaker labels.

A word on recording calls legally

Whether you may record a call depends on where you and the other person are. In the US, federal law and most states let one person in the conversation record it, while a minority of states require everyone's consent, and for a call between states it can be unclear which rule applies. Other countries set their own rules. Saying "I'm recording this so I have notes" at the start is the simple way to meet the stricter rule. This is general information, not legal advice; the guide on whether it is legal to record goes into more detail.

Transcribe your call recording