How to transcribe audio to text

Published · Updated

Three methods, and the question that decides between them before accuracy ever comes into it.

There are three ways to turn an audio file into text: upload it to an online service, use what is already built into software you own, or run the recognition on your own machine. Most guides compare them on accuracy and price. The question worth asking first is different — is this recording something you are willing to hand to someone else’s server? For an interview, a client call or a consultation, that answer decides the method before accuracy comes into it.

The question worth asking first

Almost every page on this subject opens by comparing tools. That order only works if the file is unremarkable: a podcast, a lecture, a webinar you would publish anyway.

The request is often not that. It is an interview with a named respondent, a consultation, a call where a client said something they would not say in public. For that file the comparison of accuracy scores is beside the point, because many of the pages answering this question begin by asking you to upload it, and that step is the one you cannot take.

So: decide whether the file can leave your computer. Then pick a method.

Three ways, and where your file goes in each

MethodWhat it needsWhere the file goes
An online serviceA browser and an uploadTo the provider’s servers, and it stays there under their terms
Software you already ownNothing newDepends on the product — some do it locally, some send it on
Recognition on your own machineA setup step, sometimes a terminalNowhere. It is read where it sits

The same three, as people, an upload or an app, compared, are set out on a page of their own.

The first row is the one the search results are made of, and for most recordings it is a reasonable answer: it is quick, it needs nothing installed, and a webinar recording on someone else’s server harms nobody. One such service, which also sells transcripts made by people, compared with an app on the Mac.

The second row is worth checking before you install anything, because you may already be paying for it. What you cannot assume is that built-in means local — several products transcribe by sending the audio to their own service, and the only way to know is to read what they say they do. The same split holds for speech to text tools in general, and what each one sends.

The third row is real and it is free. Open speech models run on ordinary hardware. The cost is not money, it is an afternoon: you install a runtime, fetch a model, and drive it from a terminal. If that sentence sounded fine to you, this is the best answer on the page and you can stop reading. If it did not, keep going.

If the recording is sensitive

This is the part the comparison articles skip, so here it is plainly, method by method.

Upload to a service. The audio is on their infrastructure. The transcript is too. How long both are kept, who can see them, and whether either is used to improve their models are three separate questions with three separate answers, and they are answered in the terms rather than on the landing page.

Software you already own. Unknown until you check. The word to look for is not “secure” but “where” — on this machine, or on ours.

On your own machine. The file is read where it sits, and no upload takes place at any point.

There is a fourth thing worth separating, and it catches people out: the audio and the text are not the same object and rarely have the same fate. A method can be honest about never moving the sound and still put the resulting transcript on a server — including ours, and this page would be worth little if it were coy about that. Ask the question for both halves.

Length, formats and the quality of the recording

Formats matter less than people expect. What a phone, a meeting app or a voice recorder produces is handled by essentially everything — and a tool worth using decides by looking inside the file rather than at the letters after the dot. Ours checks the type and whether there is an audio track at all, which is also why a video file is not a separate case: the picture is ignored and the sound is read. There is a separate page on what changes when the file is a video.

Length matters more, and it is the thing that goes wrong. If a service caps the length of a single file, find that cap before you upload rather than after the upload fails. A two-hour interview is precisely the file worth transcribing and precisely the one that hits the ceiling. With ours, how large the file is does not enter into it — there is nothing to send, so size stops being the question it is everywhere else.

And the thing that changes the result most is none of the above. A clean recording transcribed by a mediocre model beats a noisy one transcribed by the best available. If you have any control over how the audio is made — a microphone closer than the far side of a table, a room without a second conversation in it — spend the effort there. It pays more than any switch between tools.

What to do with the text afterwards

A raw transcript is not a document. It is a wall of speech with no paragraphs, speakers labelled by number if at all, and every false start faithfully preserved. Reading it end to end is slower than listening to the original.

What people actually want is one of three things: a specific passage, a summary, or a quotable record of who said what. Getting there means the transcript has to be searchable and attributed, which is a different job from producing it — and it is the job worth choosing a tool for.

Doing this on a Mac without the upload step

What we make does the recognition on your Mac: it makes a working copy, deletes it once the text is ready (if processing fails, it stays on your Mac), and never changes the original; by default the audio is not sent anywhere. The transcript itself syncs to our servers so it can be turned into something you can read. That is the honest shape of it: the sound stays, the text goes. If the sound is the part that cannot travel, this closes the case; if the text is also the part that cannot travel, no tool described on this page is right for you and we would rather say so.

When you bring the file in, you say whether it is a dictation or a meeting with several voices. Bringing your own file in is part of the free plan rather than a paid extra, as are recording, transcription and search, and none of it asks for a card. If you already have the file, there are three steps from a file to text in the app, and the other jobs the same app takes on are listed in what the free plan does, task by task. What happens to the text after that is set out on privacy.

If your files come from calls rather than from a recorder, the guides for Zoom, Teams and Google Meet start a step earlier. What works with no connection at all is on offline transcription, and the mechanics of the recognition itself are on how it works.

Questions

Can I transcribe audio without uploading it to a server?
Yes, and there are two routes. Open speech models run on your own machine and cost nothing, but installing them means working in a terminal. The other route is an application that does the recognition locally and hands you the text; ours does that on a Mac, and the text it produces then syncs to our servers. Which of those trade-offs you accept is the actual choice — the audio file staying put is common to both.
What is the most accurate way to transcribe audio?
Not the answer most pages give, because most pages are selling a tool. The gap between the leading speech models is smaller than the gap between a clean recording and a noisy one. A phone lying on a table in a room with two conversations going will defeat any of them, and a decent microphone will make all of them look good. Spend the effort on the recording before you spend it on comparing engines.
How long does it take to transcribe an hour of audio?
On a current machine, far less time than the recording itself — you are waiting minutes, not hours, and often much less than that. Online services add the upload to whatever their processing takes, which for a large file over a slow connection can be the longer half of the wait. That is the part people forget when they compare speeds.
Can I transcribe audio for free?
Yes. Open models are free and run on your machine if you are willing to set them up, and by default no audio leaves it. Most online services give you a handful of free minutes and charge after that. Our free plan covers recording, transcription and search with no card asked for, and the transcript it produces syncs to our servers. If you choose server recognition in Settings, the audio is sent to Dalm and our speech recognition providers and deleted once the text is ready.
What audio formats can be transcribed?
Most tools take the common ones — what a phone recorder, a meeting app or a voice memo produces. Asking about extensions is usually the wrong question: what decides is whether the file holds an audio track, which is why a video file is not a separate case. Length is the limit worth checking before you start rather than after.
Is it legal to transcribe a recorded conversation?
Transcribing a recording you already hold is usually a separate question from whether you were allowed to make it, and the second one is where the rules differ by country. The details are on our page about recording consent laws.
Can I transcribe several files at once?
One at a time. The file window takes a single file, and if you drag several onto the app, only the first is taken; a folder is not accepted. You don't have to wait for the first to finish: bring the next file in as a meeting and it waits its turn behind the one being transcribed.