Key Takeaways
Before you transcribe a recorded call, be sure the recording was lawful where the parties were. Jurisdictions differ on whether everyone had to consent or only one participant, and that question was settled before you pressed record. After that it is a technical problem. Call audio is narrowband, so the codec discards detail transcription depends on, and many recordings mix both sides into one channel, which makes speaker attribution shaky. Numbers, spellings and addresses suffer most. Check those against the audio rather than trusting the text.
You have a file. Someone on it said something that matters — a figure, a date, a name, a commitment — and you want it as text so you can point at it later. Turning the file into text is the easy part. The two things that decide whether you end up with something you can rely on both happened before you opened it: whether the recording was allowed, and how the audio was captured.
Take the legal question first. It doesn’t depend on anything technical, and no amount of careful transcription rescues a recording you shouldn’t have made.
Whether the recording was lawful comes first
Whether a phone call may be recorded depends on where the parties to the call were. Jurisdictions differ on this. Some require that everyone on the call consents. Others require only that one participant does, which can mean the person doing the recording. These are genuinely different rules, and which one applies to you is not something you can work out from the audio.
This is a legal question, not a technical one. That matters more than it sounds, because it means the answer was fixed at the moment the recording was made. Nothing you do afterwards changes it. Transcribing the file doesn’t change it. Deleting part of the file doesn’t change it. Getting the other person to agree later doesn’t retroactively make the original capture something it wasn’t.
If the two of you were not in the same place, more than one set of rules may be in play, and that is exactly the situation where it’s worth asking someone who actually knows the law where you are rather than guessing from something you read. It’s a short conversation and it happens before you build anything on top of the recording.
None of this is a reason to abandon a file you already have. Plenty of call recordings are made with everyone’s knowledge, or in places where one participant’s consent is enough, or by the person who was actually on the call for their own reference. But settle the question before you circulate the transcript, quote from it, or attach it to anything. A transcript travels much further than an audio file does, and it travels faster.
Why call audio is harder than a recording made in a room
A phone call is narrowband. The codec that carries the call throws away a large part of the sound before anything is written to a file, because it only needs to keep enough for a human to follow a conversation in real time. Humans are very good at filling in gaps from context. Transcription is less good at it, and the detail it would use to distinguish one sound from a similar one is part of what got discarded.
This is why the same speaker, saying the same sentence, transcribes more accurately from a recorder sitting on the table than from a phone line. It isn’t a question of the software being better in one case than the other. The information simply isn’t in the call file. Nothing downstream restores it — not noise reduction, not volume normalization, not running the file through a second tool. You are working with what survived the codec.
Add the usual conditions of a real call and it gets thinner. One person is walking. One person has a poor connection and the audio drops for a syllable at a time. Someone talks over someone else. On a room recording, overlapping speech is difficult; on a compressed call, it’s often unrecoverable, and the transcript will quietly pick one voice and drop the other rather than telling you it struggled.
Recording a speakerphone with a second device
Putting the call on speaker and recording it with another phone or a handheld recorder feels like a workaround. It’s worse than the original. The far side of the call has already been compressed down to narrowband, and now you’re capturing that compressed audio played through a small speaker, bouncing off the walls, along with the room’s echo and whatever else is in the room. You keep every loss the codec introduced and add a new layer on top of it.
If a device or service can record the call directly, use that. The speakerphone route is a last resort, and worth knowing is a last resort before you rely on it for anything precise.
One channel or two decides whether you get speaker labels
Many call recordings mix both sides of the conversation into a single channel. Everything ends up in one stream, and the clearest signal for telling the two people apart — which side of the line each one was on — is gone. What’s left is the voices themselves: pitch, pace, timbre. That works better with two dissimilar voices than with two similar ones, and it works worse on narrowband audio, because the voice characteristics that would distinguish them are exactly the kind of detail the codec discarded.
Where the two sides are kept as separate channels, attribution becomes far more reliable. Each channel is one person. There’s no inference involved. If the recording tool you use offers a choice, this is the setting worth finding, and it’s worth finding before the call rather than after.
When you’re working with a mixed recording, read the speaker labels as a helpful guess rather than a record. They’ll usually be right across a long uninterrupted stretch and wrong around the switches, especially where the two people interrupt each other. If the point of the transcript is who agreed to what, the switches are precisely where you need it to be correct, so spot-check those moments against the audio.
Getting the text out of the file
Once a call recording has been exported, it stops being a call. It’s an ordinary audio file and it can be transcribed like any other. The export step varies: recording apps, phone systems and carriers all differ, and the option is usually somewhere under a share button or a three-dot menu on the recording itself. Get the file out first, then stop thinking about it as a phone call and start thinking about it as audio.
From there the process is the same as for any audio file you want as text. SozAI transcribes recordings on iOS, Android and macOS with speaker labels, summaries and translation across 99+ languages; you can get the app here, and the same route handles voicemail messages, which are narrowband for the same reason and behave the same way. It’s one option among several, and the choice matters less than the quality of the file you feed it.
If this isn’t one call but a folder of them accumulated over months, the problem changes shape. Transcribing everything so you can find one sentence is usually the wrong instinct — the more useful goal is being able to search a backlog of call recordings rather than reading it.
If the recording might become evidence
A recording made so you can remind yourself what was said and a recording used as evidence are held to different standards. For personal reference, a transcript with a few rough edges is fine; you know what happened and the text is a memory aid. For anything formal, the original file usually needs to be preserved unedited.
That means keeping the file exactly as it came off the device. Don’t trim the silence at the start. Don’t cut out the irrelevant middle. Don’t re-export it in a different format to make it smaller, and don’t run it through a tool that normalizes the volume and writes a new file. Any of those produces a derivative, and the original is the thing that carries weight. Work from a copy and leave the original alone.
The transcript, in that setting, is a working document that sits alongside the recording rather than replacing it. It helps people find the relevant thirty seconds. The recording is what the relevant thirty seconds actually says. Whether a recording is admissible at all is a separate question, and it belongs to the same person you asked about the legality of making it.
What you will still have to fix by hand
The parts of a call people most need afterwards are numbers, spellings and addresses — a reference number, an amount, how a surname is spelled, which street. Those are also the parts narrowband audio damages most, because they carry no context. A whole sentence can be reconstructed from the words around it. A seven-digit number can’t. If one digit is wrong, nothing in the surrounding audio flags it, and the transcript will present it with the same flat confidence as everything else.
So treat every number, name and address in a call transcript as something to verify rather than something to use. Find it in the text, jump to that moment in the audio, listen. It takes a minute per item and it’s the difference between a transcript you can act on and one that’s merely plausible. If a detail matters enough that being wrong about it would cost you something, that’s the test for whether it needs checking.
Set your expectations for the rest of it accordingly. A transcript of a phone call is generally a good record of substance — the shape of the conversation, what was proposed, what was accepted, the tone of the disagreement. It is a weaker record of precise detail, and it gets weaker still where two people talk at once or where a connection dropped. Knowing which parts to trust is most of the skill. The transcript gives you a fast, searchable version of the conversation and a way back into the audio at the exact second you need. It doesn’t replace the audio, and on a narrowband file it was never going to.
