What you’ll be able to do

You’ll take a voice memo, an interview, a lecture, whatever, and turn it into text you can actually search, edit and skim. No typing. You just feed the audio in, and out comes a transcript you can clean up in minutes rather than hours.

That’s it, really. Once you’ve done it once, you’ll wonder why you ever typed up notes by hand.

Before you start

A few things worth sorting before you touch any tool.

First, consent. If you’re recording an interview, a call, a meeting, anyone else’s voice, tell them you’re recording and what you’ll do with it. This isn’t just good manners, it’s often a legal requirement depending on where you are and what the recording’s for. Don’t skip this bit.

Second, think about where the recording goes. If it’s a work call with sensitive details, a doctor’s appointment, anything personal or confidential, you probably don’t want to upload it to a cloud service. More on that later, but keep it in mind now.

Third, audio quality. This matters more than people expect. A transcript is only as good as what the AI can actually hear, so:

  • Keep the phone or recorder close to whoever’s speaking.
  • Turn off fans, aircon, background music if you can.
  • One person talking at a time helps enormously, overlapping speech is where transcripts fall apart.

Good audio in, good text out. Simple as that, but easy to forget when you’re rushing into a meeting.

Lastly, know roughly what file you’ll end up with. Voice memos are usually.m4a, phone recordings often.mp3 or.wav. Most tools handle these fine, but it’s worth knowing what you’ve got before you start uploading things.

Get set up

There isn’t one single “best” tool here, it really depends on what you’re recording and how private it needs to be.

If you want something that runs entirely on your own machine, Whisper is the one to know. It’s an open-source model made by OpenAI, it’s free, and it does a genuinely good job of transcription without ever sending your audio anywhere. For Mac users, MacWhisper wraps Whisper in an actual app with a nice interface, so you don’t need to touch any code. This is the option I’d reach for with anything sensitive, an interview with someone who’s asked for privacy, a therapy session, whatever.

If you’d rather upload and let a service handle it, Otter.ai and Descript both do this well. You upload your file, wait a bit, get a transcript back, often with some speaker separation built in. Handy if you’re not fussed about the file leaving your device.

If you’re on a Pixel phone, Google’s Recorder app transcribes live as you record, no upload needed afterwards, it just does it. iPhone users on recent iOS versions get something similar through Voice Memos, which can generate a transcript straight from the app.

And if you’re already paying for Microsoft 365, Word can transcribe audio you upload directly into the document. Useful if you live in the Microsoft ecosystem anyway and don’t want another app.

None of these need any AI experience to use. You’re either dragging a file into a box or pressing record. That’s the whole barrier to entry.

Try it yourself

Right, let’s actually do it.

If you’re using a chat assistant like ChatGPT, Claude or Gemini alongside your transcription tool, the trick is using the assistant after you’ve got a raw transcript, to clean it up. Here’s how that looks in practice.

Start by getting the raw text out of Whisper, Otter, Descript, whichever tool you picked. Then paste that raw transcript into your AI assistant and ask it to tidy things up:

Here's a raw transcript from an interview. Please clean it up by removing filler words (um, uh, you know), fixing obvious punctuation errors, and breaking it into paragraphs where the topic changes. Don't change any of the actual meaning or wording beyond that.

[paste transcript here]

If there were multiple people speaking and your tool didn’t label them properly, you can often fix that with the assistant too, as long as you tell it who said what based on context:

This transcript is from a conversation between two people: Sarah (asking questions) and Tom (answering). The speaker labels are missing or wrong. Based on the context and turn-taking, please add "Sarah:" and "Tom:" labels before each person's lines.

[paste transcript here]

And if there are names, places or technical terms the transcription tool got wrong, which happens constantly with anything niche, give the assistant a list to correct against:

Here's a transcript with some names and terms transcribed incorrectly. The correct spellings are: Aoife Ní Bhriain, Kilkenny, and the drug name "levothyroxine". Please find and correct any misspelled versions of these in the text below, leaving everything else unchanged.

[paste transcript here]

Once it’s clean, you can ask for a summary too:

Please summarise this transcript in about 150 words, covering the main points discussed and any decisions or action items mentioned.

[paste transcript here]

That’s four separate jobs, cleaning, labelling, correcting names, summarising, and you can run them one after another or combine a couple if the transcript’s short enough.

Check the result

Don’t just trust the output blindly. A few checks worth running every time.

Read the first minute against the actual audio. Listen while you read. If it’s roughly right there, it’s usually roughly right throughout, but this catches early problems before you’ve wasted time on the rest.

Check names specifically. Transcription tools are notoriously bad with proper nouns, especially names that aren’t common in English. Scan for anything that looks garbled or oddly spelled, then compare against a list if you have one.

Look for speaker mix-ups. If two people have similar voices or talk over each other, tools sometimes attribute a line to the wrong person. This matters a lot in interviews, so check any quote you plan to use publicly.

Search for numbers and dates. These get mangled more than people expect, “fifteen” becoming “50,” that sort of thing. Worth a quick scan if the recording had any figures in it.

And check the ending. Some tools cut off the last few seconds if the file didn’t upload cleanly. Compare the transcript’s last line against your actual recording’s last line.

None of this takes long. Five minutes of checking against ten minutes of audio is a fair trade for not embarrassing yourself with a misquote.

If it doesn’t work

A few common problems and what’s usually behind them.

If the transcript is a mess of gibberish or huge gaps, it’s almost always audio quality. Background noise, overlapping speakers, a phone that was too far away, these all confuse the tool badly. There’s not much you can do after the fact except try a different tool and see if it copes better, but really the fix is prevention, better mic placement next time.

If the tool won’t accept your file at all, check the format. Most tools handle.mp3,.wav and.m4a fine, but some struggle with less common formats or very large files. Converting to.mp3 with a free converter usually solves this.

If speaker labels are wrong or missing entirely, that’s often just a limitation of the tool rather than a fixable bug, some are better at this than others, and it’s part of why I mentioned checking this manually above.

Now, the important bit: privacy.

Don’t upload anything you wouldn’t want sitting on someone else’s server. That includes medical information, legal conversations, anything covered by a confidentiality agreement, and honestly anything involving another person who hasn’t agreed to it being processed by a third-party tool. If in doubt, use Whisper locally, it never leaves your machine, which makes it the right call for interviews with vulnerable people, sensitive work discussions, or anything you’d rather not think about ending up somewhere you didn’t intend.

And always get consent before recording someone else. Not just because it’s the right thing to do, but because handing an AI tool a recording of someone who didn’t know they were being recorded is a genuinely bad idea, ethically and often legally.

Keep exploring

If this was useful, a few related guides worth a look: how to get AI to take meeting notes, how to get AI to summarise a YouTube video, and how to get AI to read text out loud. Different job, same basic idea, feed the AI something and let it do the tedious part.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.