About TapeText
Who makes it
Damien Yvert, an independent developer. TapeText is one of a set of small browser tools I build and run on my own. The others compress video, cull RAW photographs, draw charts and make résumés. They have one thing in common, which is the reason I keep making them: each does something useful without the file ever leaving the machine it is on.
Why the model runs in your browser
Transcription is the clearest case I have found of a task where the ordinary architecture is wrong for a large minority of the people who need it. Sending audio to a server is cheaper, faster and easier to build, and it is unavailable to a journalist with a source, a lawyer with a deposition, a doctor with a consultation, a researcher with a consent form that named who would hear the recording.
Those people are told to install something instead, which assumes they can install software on the machine in front of them. Often they cannot. Running the model in the browser is the only arrangement that asks for neither the upload nor the install.
What it costs
I would rather set this out than have you discover it. Running a speech model on your own machine means a download of 41 to 146 MB the first time, a wait measured in minutes rather than seconds, and a hard requirement for a reasonably recent computer. A phone will not do it well, and the page says so before anything starts rather than after.
It also means the model is Whisper in its small and base sizes rather than the largest one, because the largest is a gigabyte and nobody is downloading that to read one interview. Accuracy is good and it is not perfect; the editor exists because of that, not in spite of it.
If none of those trade-offs suits you, a server-side service will be faster and more accurate, and you should use one. The page on not uploading is about the situation where that is not an option.
How it is paid for
It is not, at the moment. There is no server doing the transcription, so the tool costs almost nothing to run, which is why there is no account, no cap and no trial. Everything it does today stays free.
A paid add-on is planned later: processing a folder of files in one go, a larger and more accurate model, formatted document export, and a vocabulary of names and jargon that a general model mishears. It will be a one-time purchase rather than a subscription, and it will not take anything away from what is free now.