1
The need
Turning a recording into text you can use: the minutes of an interview, a document for a case file, a corpus to analyse, subtitles for a video. Without spending the day on it.
Transcription and subtitles for recordings
Your recordings, transcribed on your own computer
Klarword turns an interview, a meeting or a video into timestamped text: you read it back while listening, correct it, and export it to Word or as subtitles. A video comes back to you subtitled, ready to publish. It is built for professions that record people, and the work is done in your browser: the audio stays on your computer.
What Klarword is for
An hour of interview usually takes four to six hours to type up by hand. Online transcription services are much faster, but you have to hand them the recording.
2
The catch
An online service asks you to upload the audio to it. For an HR interview, a consultation or a source, that means personal data held by a third party: one more processor, and a retention period that is no longer up to you.
3
The answer
The same automatic transcription, but the model runs in your browser. The text arrives in a few minutes, and the recording never leaves your computer.
What it looks like
An excerpt from an interview, as Klarword renders it: in the editor, as subtitles and as a Word document.
1
00:00:10,000 --> 00:00:15,000
Could you describe your career here since
you joined in 2019?
2
00:00:15,000 --> 00:00:22,515
I started in the accounts team in Leeds,
then I moved to the HR team in Wakefield
3
00:00:22,515 --> 00:00:24,000
three years ago.What Klarword does
Your files as they are
MP3, M4A, WAV, MP4 and most formats your browser can play. Or record from the microphone and watch the text appear as people speak, or send the file from your phone with a QR code.
A timestamped transcript
English, French and about fifteen other languages. Each segment carries its time, and the text appears as the computation goes.
An editor that follows playback
Each word carries the moment it was spoken: Ctrl+click to hear it, split at the exact word. Words the model was unsure of are highlighted, Alt+U takes you to the next one, Ctrl+Z undoes. Tags and notes to code an interview.
The exports you need
Word in four layouts, PDF, text, and subtitles in SRT, WebVTT, TTML, SBV or ASS, checked against broadcasters' standards. Raw verbatim or cleaned of hesitations. No author, no date.
Your video, returned with subtitles
Klarword gives you back the video itself, subtitles burned in: outline, banner, social media or high legibility, one colour per voice, bilingual with the translation. Faces blurred if needed, a vertical format that follows whoever is speaking, and one clip per chapter for social media.
An explicit choice of model
Fast, Balanced or Accurate, with the exact download size for your computer, shown before you start, and a measured speed.
Minutes written on the spot
Summary, key points, decisions and actions, then timestamped chapters and quotes checked word for word, by a small language model that runs on your graphics card. A draft, which never leaves the computer.
Subtitles translated on the spot
The same language model translates the lines into English, German, Spanish and other languages, at the original timings. Nothing goes to an online translator.
Who speaks, identified
Voices are identified on your computer by two small models, then you rename each one once for the whole transcript.
Your own glossary
Proper names and professional vocabulary, given to the model before transcription so it spells them right from the start. Whatever still slips through is corrected automatically or suggested to you.
Pseudonymise before sharing
Proper names are found and suggested; the ones you tick become “Person A”, “Person B”, in the text and in the speaker names alike.
A batch, or yesterday's work
Drop ten interviews, get back a ZIP. Keep a piece of work on this computer to pick it up tomorrow, search for a word across all your work, pass it to a colleague in a .klarword file, or review it together, each on your own computer.
For recordings that must not travel
Interviews, hearings, consultations, fieldwork: the professions where people are recorded, and where someone answers for it.
Human resourcesHR interviews that stay inside the company
Lawyers and legal teamsPrivilege, all the way to the transcript
HealthcareHealth data that does not leave the practice
Social science researchQualitative interviews you do not have to hand over
JournalismProtecting sources starts with the transcript
Public sector and local governmentMeeting minutes without an outside provider
Three models, a stated size
All three models come from OpenAI's Whisper family, released under the MIT licence. The size shown is what your computer will download, once; it depends on the graphics card.
| Model | What it is for | With a graphics card | On the processor |
|---|---|---|---|
| Fast74 million parameters | For a clean recording, a quick read-through or a modest computer. | 80 MB | 80 MB |
| Balanced244 million parameters | The right trade-off for an interview on a recent desktop computer. | 302 MB | 252 MB |
| Accurate809 million parameters | The best quality: proper names, accents, overlapping voices. Needs a graphics card. | 566 MB to 762 MB | 762 MB |
Add, once, the compute engine shared by all three: 27 MB.
Measured speed
Computation time for one hour of recording, measured on a recent laptop: AMD Ryzen 9 6900HX (8 cores), 16 GB of memory, integrated Radeon 680M graphics. A desktop computer with a dedicated graphics card is faster; an older computer, slower.
| Model | With the graphics card | On the processor alone |
|---|---|---|
| Fast | on the processor | about 13 min |
| Balanced | about 19 min | about 27 min |
| Accurate | about 19 min | about 104 min |
Measured with Chrome 152, on a 71 s interview scaled up to an hour; model already stored on the computer, download not counted.
No installation, and no upload
Other solutions ask you either to send the audio or to install software.
| Online services | Installed software | Klarword | |
|---|---|---|---|
| The audio leaves your computer | Yes, to the provider's servers | No | No |
| Installation and administrator rights | No | Yes, often blocked on company computers | No, an address is enough |
| Works with the network off | No | Yes | Yes, once the model is loaded |
| What you have to trust | Terms of service and subcontractors | The software publisher | Nothing: it can be checked in the browser |
What Klarword does not do
- It does not recognise voices with certainty. Speaker identification gets confused by two similar voices and very short interruptions: check it, and split at the cursor any segment that mixes two people.
- It does not guarantee an exact transcript. A model makes mistakes on proper names, overlapping voices and noisy recordings: proofread before you share.
- It keeps nothing once the tab is closed. Save your work; a WebVTT export reopens in Klarword to resume proofreading.
- It does not replace informing the people you record. Consent remains your responsibility, as the terms of use point out.
Try it with a real recording
The Fast model weighs under 80 MB. Choose a file, disconnect the network once the model is ready, and watch.
Transcribe a file