Transcribe audio into something you can search

You have an hour of audio — an interview, a lecture, a meeting you recorded — and you need one thing someone said in the middle of it. Listening again to find it costs an hour. A transcript costs minutes and turns the recording into something you can search, quote and skim.

To transcribe audio means converting speech into text, and there are three ways to do it that differ enormously in cost, privacy and quality. Choosing between them is straightforward once you know which questions decide it.

The three routes

Your device already does this. Both phone platforms have dictation and live transcription, and recent versions can transcribe an existing recording rather than only live speech. It is free, it runs locally on current hardware, and nothing is uploaded — which makes it the correct default for anything sensitive.

A dedicated tool buys three things the built-in option generally lacks: handling long files without babysitting, speaker labels, and timestamps linking text back to the audio. Recorders that transcribe on the device — TapMemo among them — combine the two, keeping the audio local while producing a searchable transcript.

By hand is four to six times the length of the recording, and it is the right answer only for a short clip where every word must be exact.

Decide privacy first. If the audio should not leave your device, that eliminates most of the options before quality enters the conversation.

What decides the quality

When you transcribe audio, the tool is rarely the deciding factor. Four things are, in order of impact.

• Microphone distance. A phone on the table three metres from the speaker produces a transcript no software can rescue. Closer is the single largest improvement available and it costs nothing.
• Background noise. Continuous noise — air conditioning, traffic, a fan — degrades recognition badly, and it is worse than occasional loud sounds because it never lets the model recover.
• Overlapping speech. Two people talking at once is close to unrecoverable. Meetings where people interrupt transcribe far worse than interviews where they take turns.
• Vocabulary. Names, technical terms and acronyms are what a general model has least reason to know, and they are usually the words that matter most.

That ordering has a practical consequence: spend your effort on the recording, not on comparing tools. A good recording transcribed by a mediocre tool beats a poor recording transcribed by the best one available.

Recording so it transcribes well

Five minutes of preparation before you record, and the difference when you transcribe audio afterwards is not subtle.

1. Put the microphone near the speaker. A phone at arm's length, or lapel microphones if you have them. In a meeting, the device goes in the middle of the table, not beside one person.
2. Kill continuous noise. Turn off the air conditioning, close the window, move away from the fridge.
3. Choose a soft room. Hard surfaces produce reverb that smears the sound; a room with a rug, curtains or soft furniture is dramatically better.
4. Ask people to take turns, once, at the start. It feels awkward for ten seconds and doubles the usable transcript.
5. Say names at the beginning. Both a courtesy and a hint to the model, and it makes the transcript far easier to attribute afterwards.

Point one outweighs the rest combined. Everything else is refinement on top of getting close.

Cleaning up what comes back

However you transcribe audio automatically, the result needs a pass. Five minutes of editing produces something usable for months.

Search for the names first. Product names, people, places. This is where the errors cluster and where they matter most, and a find-and-replace fixes each one everywhere at once.

Add paragraph breaks. Automatic transcripts arrive as walls of text. Breaking at topic changes is what makes it skimmable, and skimmable is the whole point.

Mark the speakers if the tool did not, at least at each change of speaker.

Leave the filler in if you are quoting. Cleaning up "um" and false starts is right for notes and wrong for a quotation you attribute to someone.

Keep the timestamps. They are what lets you return to the audio for anything ambiguous, and deleting them removes the transcript's link to its source.

The last one is the habit most worth forming. A transcript with timestamps is a searchable index of the recording; one without is just text that may or may not be accurate.

Long recordings, and where they go wrong

An hour of audio behaves differently from five minutes, and three problems appear only at length.

The process gets interrupted. A local transcription of a long file is minutes of sustained work, and the phone locking, a call arriving or the app being backgrounded can stop it. Keep the screen on and the device charging for anything over about twenty minutes.

Quality drifts. Speakers move, the room changes, batteries in a lapel microphone fade. A transcript that is excellent for ten minutes and poor thereafter usually reflects the recording rather than the software.

The output becomes unwieldy. An hour of speech is roughly 8,000 to 10,000 words, which is a document rather than a note. Without headings, timestamps and a summary it is technically searchable and practically unusable.

The fix for all three is to split long recordings at natural boundaries — topics, agenda items, sides of a conversation — and transcribe each piece separately. Smaller pieces survive interruption, make quality problems local rather than global, and produce a document with structure already in it.

One more thing that only shows up at length: check the first and last minute specifically. Recordings frequently start before people settle and continue after the useful part ends, and trimming those before transcribing saves cleaning them out of the text afterwards.

Making it actually searchable

A transcript in a note is better than an hour of audio, and there are three cheap steps beyond that.

Name the file properly. Date, subject, participants. This is the difference between finding it in six months and not.

Put a summary at the top. Three lines: what this was, who was there, what was decided. Most of the time you will read only that.

Store it as text, not as an image or a PDF of text. Plain text and Markdown are searchable by everything, forever, on every device — which is also why they survive tool changes. A reader like Mdora is a quick way to check that a file really is text rather than a picture of it: if the words select, they are searchable.

For a recurring series — a weekly meeting, a podcast, a course — a consistent filename and a consistent summary format turns a folder of transcripts into something you can actually search across, which no individual transcript achieves on its own.

When not to transcribe

Two cases where it is wrong to transcribe audio at all.

When you need the tone. A transcript loses hesitation, emphasis and irony. For anything where how it was said matters, keep the audio and use the transcript only as an index.

When you cannot legally hold the recording. Consent rules for recording conversations vary by country and sometimes within one, and a transcript is a record of the same content. Ask before recording; it takes three seconds and removes the whole problem.

More text tools in text tools, the recording side in video tools, and general picks in online tools. Apple documents its on-device dictation and transcription if you want the free route first.

The short version

To transcribe audio, start with what your device already does — free, local, and good enough for clear speech. Move to a dedicated tool when you need long files, speaker labels or timestamps, and decide the privacy question before the quality one.

Quality is decided by the recording rather than the software: get the microphone close, kill continuous noise, ask people to take turns. Then spend five minutes fixing names, adding paragraph breaks and keeping the timestamps — because a transcript you can search back to the audio is worth far more than a slightly more accurate one you cannot.