Frequently asked questions
Is my recording sent anywhere?
No. Your browser reads the file and transcribes it on your computer. The content security policy forbids the page any connection to another domain, and this site's server has no address that could receive a file. The Proof page explains how to check.
So what is downloaded?
The pages, the code and, the first time, the compute engine and the chosen model, all from this site. The model is then stored on your computer and does not download again.
Does it work without a connection?
Yes, once the model is loaded. You can disconnect before dropping your file, and even reload the page: it comes back from your browser's cache.
What computer do I need?
A recent computer with 8 GB of memory and Chrome or Edge gives good results. A graphics card the browser can use (WebGPU) makes the Balanced and Accurate models much faster. Chrome and Edge are the recommended browsers; on Firefox and Safari, it is not guaranteed to work.
What about a phone?
The Fast model runs there, slowly. Klarword is designed for a workstation: heavy models, intensive computation, proofreading at the keyboard. On mobile, the site tells you so and offers to open the address on your computer.
Which formats can I import?
MP3, M4A, WAV and MP4, and most formats your browser can play (OGG, WebM, FLAC, MOV depending on the system). The sound track of a video is extracted on the spot.
Which languages are recognised?
English, French and about fifteen other languages, with automatic detection. Quality is better with the Accurate model, especially outside English.
How long does a transcription take?
It depends on the model and the computer. The tool shows the actual speed while it works, as a multiple of real time, and the estimated time left.
Is the transcript accurate?
It is good on a clean recording, and fallible on proper names, overlapping voices and noisy rooms. Proofread it before sharing: that is what the editor is for.
Does Klarword recognise who is speaking?
Yes, on your computer: once transcription has finished, “Identify speakers” finds the voices (automatically or for a given number), then you rename each one just once. Two small models totalling 33 MB do the work, on the processor. Exports use the names: paragraphs per person in Word, voice tags in WebVTT. Check the result: two similar voices can be mixed up.
Is my work saved?
No, nothing is kept when the tab closes, by design. The browser warns you if you leave without saving. A WebVTT export reopens in Klarword, with the recording, to resume proofreading.
Where are the models stored, and how do I delete them?
In the private space the browser keeps for this site. The tool's “On this computer” panel lists and deletes them; clearing the site's data in your browser settings erases them too.
Who must obtain the consent of the people recorded?
You do. Klarword is a tool; informing people and obtaining their agreement remains the responsibility of whoever records. See the terms of use.
Can we install it on our internal network?
Yes: that is self-hosting. The whole site, models included, sits on your intranet and works with no outside connection at all.