Interview transcription

Interviews are often the most sensitive data a project holds: names, health, work, family. Cloud transcription services receive the recording, keep it on their servers and process it under their own terms, which is why many universities and companies tell staff not to upload interviews to them.

This tool sends nothing. OpenAI's open Whisper model and the speaker detection run inside your browser, on your own processor and graphics chip. The only download is the model itself, the first time. After that the page works with Wi-Fi switched off, which is the simplest way to check the claim yourself. The data-flow statement lists every request the page makes.

We tested it on three recorded hour-long interviews with professional reference transcripts. Accurate (Whisper large-v3-turbo) got 8.3% of words wrong, on a par with desktop Whisper tools run on the same files. Most of the difference from the reference is "um", "uh" and repeated words, which Whisper leaves out.

Speaker labels held for the whole hour: fewer than 2 words in 100 were given to the wrong speaker. Telling the tool how many people speak helps; two is the right answer for most interviews.

• Record close to each speaker. A phone in the middle of a table works; a phone in a pocket doesn't.
• Quiet rooms beat expensive microphones. If the recording is noisy, run it through the background noise remover first: cleaner audio means fewer wrong words.
• Pick the language if you know it. Detection reads the first 30 seconds, and a greeting in another language can mislead it.
• Check names, places and technical terms. Whisper spells them the way they sound; click the line, listen, and fix it in place.

Accurate is fastest in Chrome or Edge on a computer with WebGPU: on a 2020 laptop's built-in graphics it transcribed an hour-long interview in about an hour, while the speakers were found in parallel. Newer laptops are faster. Without WebGPU it runs on the processor, more slowly.

Fast (Whisper small) runs on the processor at about twice real time and downloads 249 MB instead of about 560 MB. 8 GB of memory is enough for a three-hour recording.