Free transcription on a Mac: what each free route costs instead

Search for free transcription software for Mac and you get two kinds of results: web converters that want your file, and paid apps with the word “free” in the meta description. Both hide the same question, which is what the free version actually costs you.

There are four honest free routes on a Mac — and they cover video transcription free of charge just as well as audio, since the audio track is what any of them reads. I ran all of them on the same recordings and measured what each one gives back.

The four routes

Route	Cost in money	Cost in something else
Apple Dictation	$0	Live speech only; no files
A free tier (MacWhisper, superwhisper)	$0	No file transcription
whisper.cpp in the terminal	$0	A terminal, and 141 MB up
An online converter	$0 up to a cap	Your file goes to their server

Two of these keep your audio on the machine and need no account. Two of them do not do the job most people arrive with — a recording that already exists.

Apple Dictation: already installed, and narrower than you think

It costs nothing, it is in every text field, and on Apple silicon Apple documents that voice inputs and transcripts for general text dictation are processed on the device rather than sent to Siri servers. As speech to text free Mac options go, it is the one with no download at all.

Its limit is not accuracy. It is that Dictation listens to you speaking now — there is no way to hand it the interview you recorded last week. What it covers, where it stops and what changed in recent macOS releases is a longer piece of its own.

Free tiers: complete products, with one line drawn

Two Mac apps give away more than most paid ones.

MacWhisper’s free tier (its pricing, what the paid tier adds) covers recording from the microphone, mp3, wav, m4a, ogg, opus, mov and mp4, 100 languages, export to SRT, VTT and TXT, and system-wide dictation.

superwhisper’s free tier (its pricing, where its paid line falls) covers voice to text in any app, meeting recording and transcription, 100+ languages and unlimited use of Whisper models running locally, with no weekly cap.

Both are genuinely free forever rather than trials. And both draw the same line in the same place: transcribing audio and video files you already have is the paid feature. If your job is a folder of recordings, the free tier of a Mac app is not the route — which is what sends most people to the terminal or to a web uploader.

whisper.cpp: free in every sense, if you can type a command

This is the route with no asterisk. It is the same Whisper model the paid apps run, on your machine, with nothing uploaded and nothing capped.

Here is what it cost, measured rather than estimated:

• 8.6 MB — the program itself, 33 files, installed with brew install whisper-cpp (whisper.cpp, MIT licence);
• 141 MB — the smallest useful model, downloaded in 39 seconds;
• 0.37 seconds — to transcribe a 26-second clip, about 70 times faster than real time.

Two minutes from nothing to a transcript. The command is one line:

whisper-cli -m ggml-base.bin -f recording.wav -nt -otxt -of transcript

That is the whole route. There is no account, no upload, and no limit on how many hours you feed it.

What free costs in accuracy

Here is the part the free-tool listicles never measure. I ran three Whisper model sizes over the same 200 words of synthesised meeting speech — product names, two surnames, half a dozen pieces of jargon — and scored them against the reference text.

Model	Size on disk	Download time	26 s clip took	Errors in 200 words	WER
base	141 MB	39 s	0.37 s	35	17.5%
small	465 MB	~2 min	0.65 s	33	16.5%
large-v3-turbo	1,549 MB	~7 min	0.99 s	23	11.5%

Three things fall out of that table.

Free is not the same as bad. Even the 141 MB model produces a readable transcript of a technical standup. Most of what it gets wrong, it gets wrong in ways you can see and fix.

The middle size is a waste of a download. base to small is 3.3 times the disk for two fewer errors — inside the noise on 200 words. The jump that pays is to a large model, which removed a third of the errors. Download the smallest or the largest; skip what is between.

Names are where all of it fails. Kekoso came back as coco, Grafana board as griffin aboard, kubectl as cube to, rollback as after all book. That is not a free-tier problem, it is what every model does with vocabulary it has never seen. Feeding the same recordings through a custom vocabulary took errors to 4.0% — a bigger improvement than any model size in that table, and one that costs nothing at all.

One caveat on my own numbers, since measuring word error rate honestly includes admitting this: the small model scored 15.0% in an earlier run and 16.5% in this one, on the same audio. Three errors of drift on 200 words. Treat differences under about two points as ties.

Online converters: the free tier you pay for in a different currency

Web converters are free up to a cap, need no install, and work on a phone. The cost is the upload: your recording sits on someone else’s server under terms you did not read, and the free tier usually ends at a file size, a minute count, or the second file.

For one short recording of nothing sensitive, that is a fair trade and I would not argue against it — our own converter page names the free online route first for exactly that case. For a client interview, a medical appointment or anything under NDA, the calculation is different and it is not about quality — what “on-device” means, and how to check a claim is the test to apply before uploading anything.

So when does paying make sense?

Free covers more than the search results suggest: dictation on Apple silicon, unlimited local dictation in two free tiers, and unlimited file transcription from the terminal. If that is your shape of work, the honest answer is that best free transcription software for Mac is a real category and you may never need to buy anything.

Paying buys three things, and none of them is a better transcript:

1. A window instead of a command. File pickers, batch queues, a transcript you can click through and edit.
2. The file half of the job in an app. Both free tiers stop exactly there, so this is the most common reason people spend money in this category.
3. The things around recognition. Speaker labels, meeting capture, vocabulary that persists, subtitles with sensible cue boundaries, an agent that can call it.

Kekoso, which I build, is $29 once for three Macs and sits in that third group — file transcription included rather than gated, calls recorded as separate tracks, and a local MCP server. It is not a better model than the free one; it is the same model with the parts around it built.

Which is the thing to hold on to when comparing anything in this category: the recognition is free. What varies is everything else.

Measured 7 September 2026 on a MacBook Pro with an M4 Max, whisper.cpp 1.9.1 from Homebrew, against reference text written before the audio was synthesised. Free tier contents read from each vendor’s own pricing page the same day.

Questions people ask

What is the best free transcription software for Mac?

It depends which cost you would rather pay. whisper.cpp from Homebrew is free in every sense — 8.6 MB of program, a model from 141 MB up, nothing uploaded — but it is a terminal command. MacWhisper's free tier and superwhisper's free tier are free apps with a graphical interface, and both stop short of transcribing files you already recorded. Apple's built-in Dictation is already installed and needs nothing.

Can I transcribe audio to text free without uploading it?

Yes. A local model does the work on your own machine with the network off. The free routes that qualify are whisper.cpp from the terminal, the free tiers of local Mac apps, and Apple's own Dictation. Online converters are the ones that require an upload, and that is the price of their free tier.

How good is free transcription, really?

On my own 200 words of meeting speech, the free 141 MB Whisper base model made 35 errors — 17.5%. The 465 MB small model made 33, and the 1,549 MB large-v3-turbo made 23, which is 11.5%. Nearly all of the errors were names and jargon. A custom vocabulary was worth more than any of those size steps, cutting errors to 4.0% on the same material.

Is a bigger Whisper model worth downloading?

Only the jump to a large model is. Going from base at 141 MB to small at 465 MB moved my result by two errors in 200 words, which is noise. Going to large-v3-turbo at 1,549 MB removed a third of the errors. So download the smallest or the largest, and skip what is in between.

Is free dictation software for Mac good enough for daily use?

For dictation into apps, yes — superwhisper's free tier includes voice to text in any app and unlimited use of Whisper models locally, with no weekly word cap, and MacWhisper's free tier includes system-wide dictation. What free tiers usually withhold is file transcription, batch handling and anything cloud-assisted.

Does Apple's built-in Dictation cost anything?

No, and on Apple silicon Macs Apple documents that general text dictation is processed on the device rather than sent to Siri servers. Its limits are elsewhere: it works on live speech rather than on files you already have, and it has no way to transcribe a recording sitting in your Downloads folder.

How long does it take to set up free local transcription?

Under two minutes on a normal connection. Homebrew installs whisper.cpp as 8.6 MB across 33 files, and the base model is a single 141 MB download that took 39 seconds here. After that a 26-second clip transcribed in 0.37 seconds — about 70 times faster than real time.