Why every free transcription site caps you at 30 minutes (the maths)

A free transcription site caps you because every minute you upload costs it money on a server, and a visitor looking at ads does not pay for more than about five of those minutes. The cap is where the two lines cross. This post does the arithmetic with published prices, then explains why a site that runs the model on your computer instead of its own can skip the cap entirely.

What a free minute costs the site

Speech recognition is sold wholesale by the minute, and the wholesale prices are public. As published on 15 September 2026, AssemblyAI lists asynchronous transcription at $0.15 an hour for Universal-2 and $0.21 an hour for Universal-3.5 Pro, which is $0.0025 and $0.0035 a minute. Deepgram lists Nova-3 pre-recorded at $0.0043 a minute for monolingual audio and $0.0052 for multilingual. OpenAI lists $0.006 a minute for Whisper and for gpt-4o-transcribe, and $0.003 for gpt-4o-mini-transcribe.

A site running its own servers pays less per minute at scale, but it pays for those servers whether anyone uploads or not, so the API price is a fair proxy for the marginal cost of one more free minute.

Take the middle figure, $0.0043 a minute. A 30-minute file costs about 13 cents to transcribe. A 300-minute monthly allowance, which is what Otter publishes for its free Basic plan, costs up to $1.29 per user per month if it is used in full.

Then there is the upload itself. An hour of MP3 at 128 kbps is about 58 MB: 128,000 bits a second times 3,600 seconds, divided by eight. The site has to receive that, hold it at least until the job finishes, and usually keep the transcript afterwards. Cloud storage and bandwidth are cheap per gigabyte, but a free user who uploads ten hours a month is half a gigabyte of traffic and storage that nobody paid for.

What a free visitor earns the site

If the site shows ads, the measure is revenue per thousand page views. Google’s AdSense help defines page RPM as estimated earnings divided by the number of page views, multiplied by 1,000.

We are not going to quote you an industry average, because the honest answer is that it depends on the niche, the country and who is bidding, and the figures that circulate are not from primary sources. Put your own number in. To show the working, take a page RPM of $8, which is 0.8 cents a page view.

Now put the two numbers next to each other. Transcribing a 30-minute file costs the site about 13 cents. At 0.8 cents a page view, that visitor has to look at roughly 16 pages to pay for it. Most people who arrive at a tool page look at one or two. Halve the RPM and it takes 32 pages. Double it and it still takes eight.

So a 30-minute cap does not make the free user profitable. It caps the loss at about 13 cents, which is a marketing cost a site can justify if a small share of free users go on to pay. Set the cap at 30 minutes a file and 300 minutes a month and the worst case is roughly $1.30 per user per month, which is still comfortable against a paid plan in the teens of dollars.

The published free limits, as of 15 September 2026

Here is what four services state on their own pricing pages. These are the limits they publish, nothing more. Why any particular company sets its number where it does is its own business, and none of them have said.

Two things are worth noticing. The allowances all sit in the same band, tens of minutes to an hour or so a month, which is what you would expect if they are all buying compute in the same market. And the limit that bites first is often not minutes but files, which is the cheaper thing for a server to count.

Why the numbers push everyone the same way

Once a transcribed minute costs real money, three things follow from the arithmetic rather than from anybody’s intentions:

1. There has to be a boundary. Per file, per month, or both. The sum above puts it roughly where you find it.
2. The boundary has to attach to someone. A limit you cannot enforce is not a limit, and enforcing one means knowing who is in front of you, which is what a signup does.
3. The boundary is where the paid plan lives. A capped free tier is a trial, and a trial is defined by where it stops.

None of this is sinister. It is what the cost structure requires. But it does mean that “free” on a site which transcribes on its own servers always carries an asterisk, and the asterisk is the cap. The cost per hour comparison works the same arithmetic through the paid tiers, where the numbers get easier to compare.

What changes when the model runs on your machine

The same speech model family the cloud is built on, OpenAI’s open-source Whisper, has been converted to run inside a web browser on the visitor’s own graphics card through WebGPU. What WebGPU actually is covers the machinery. When a site takes that route, the cost line moves.

The marginal cost per minute becomes zero. There is no upload, so no bandwidth. There is no server job, so no compute. There is no stored file, so no storage. The only thing the site serves is the model itself, about 200 MB once, which the browser then caches.

Redo the sum with a zero in it. A 30-minute file costs the site nothing. A three-hour file costs the site nothing. The visitor still earns whatever a page view earns, and now all of it is margin. There is no reason to cap, no reason to require an account, and no point in the flow where a price has to be shown. A few ads are enough to cover hosting, which is how this one is run.

The cost has not vanished. It has moved to you, as a model download and roughly 40 minutes of your computer’s time per hour of audio. For most people that is the better side of the trade, because the computer was going to be on anyway.

The trade-offs, honestly

Local transcription has its own asterisks, and they are different ones:

• Your hardware sets the speed. On our 2021 test laptop, using its integrated Intel graphics, 25 minutes of audio took about six and a half minutes, so an hour is roughly a quarter of an hour on that machine; the benchmarks page has the runs. A thin laptop takes roughly as long as the recording.
• It needs a browser with a working WebGPU adapter, which today means desktop Chrome or Edge. Without WebGPU the model runs on the processor instead, which is slower; the browser support page says which browsers do what.
• The model here is the base English Whisper, picked so it runs on ordinary hardware rather than for top accuracy. It is English only for now.
• Accuracy follows from that choice. On our clean synthetic test clip it made no word errors, and one on the same clip with background noise (the examples page shows both). Names, technical vocabulary, heavy accents and noisy rooms are the known weak spots, so read the transcript through before you rely on it.

What you get back is no cap, no account, no upload, and a recording that never left your disk. For lectures, interviews, meeting notes and first-draft subtitles that is usually the right side of the trade.

Frequently asked questions

Why not just charge less? Because the per-minute cost underneath is real and published: $0.0025 to $0.006 a minute at the four APIs above, or 15 to 36 cents an hour. A consumer product is more than the recognition, so it costs more than the recognition, and there is not much room underneath the API price to work with.

Could an ad-funded site afford no cap on a server? Only if visitors averaged around 16 page views for every 30-minute file, at an $8 page RPM. A tool page does not usually get that, because people come to do one thing and leave.

Is a browser model as accurate as a large cloud model? We would rather tell you where ours slips than rate anyone else’s. This one runs the base English Whisper model. It made no word errors on our clean synthetic clip and one on the noisy version, and it struggles with names, technical terms and heavy accents; the examples page shows the runs. Read it through before you publish anything from it.

Does “runs in the browser” mean my file is uploaded and deleted afterwards? No. Nothing is sent. The processing happens in the tab, and you can watch the browser’s network panel stay empty while it runs.

Sources, checked 15 September 2026

• https://www.assemblyai.com/pricing : AssemblyAI pricing (asynchronous speech to text at $0.15 an hour for Universal-2 and $0.21 for Universal-3.5 Pro)
• https://deepgram.com/pricing : Deepgram pricing (Nova-3 pre-recorded at $0.0043 a minute monolingual, $0.0052 multilingual)
• https://developers.openai.com/api/docs/pricing : OpenAI API pricing (Whisper and gpt-4o-transcribe at $0.006 a minute, gpt-4o-mini-transcribe at $0.003)
• https://otter.ai/pricing : Otter pricing (Basic: 300 monthly transcription minutes, 30 minutes per conversation, three lifetime file imports)
• https://www.rev.com/pricing : Rev pricing (Free: 45 AI transcription and caption minutes a month, analyse up to five files at a time)
• https://www.descript.com/pricing : Descript pricing (Free: 60 media minutes a month, 100 one-time AI credits)
• https://sonix.ai/pricing : Sonix pricing (30 minute free trial, pay as you go at $10 an hour)
• https://support.google.com/adsense/answer/190515 : Google AdSense help, the definition and formula for page RPM
• MP3 size arithmetic: 128 kbps times 3,600 seconds divided by 8 is 57.6 MB an hour
• Our own numbers for speed, model size and accuracy come from our test of the transcriber on this site