The numbers

How well Ratproof works, and what it runs on. Every number is for the version you can use today.

Accuracy

In English, about 91% of the words it writes match the human transcript. Tested on real meetings with four people talking over each other. That audio is hard on purpose. That is why researchers use it.

• About 93% with Accuracy switched on in Settings. It uses a larger model, so it takes about 1.5 times as long in Chrome or Edge on a Windows PC, and about twice as long on other computers. It needs a 240 MB extra download.
• Silence is skipped. Only the parts where someone is talking are transcribed.
• Loops are flagged. Text that gets stuck repeating is marked in your transcript and in the files you save.
• Your text is never changed. You are told, not corrected.

Speed

• First text in about 2 seconds in English on the test machine, once the model is downloaded.
• 7 to 30 minutes per hour of audio in English, depending on your computer. Text appears as it goes, so you can read along while it works.
• Full speed while you use other windows. A small floating window shows progress, and it stops Chrome from slowing the work down.
• About 235 MB on the first visit, then saved. Later visits download nothing.

Speaker labels

Press “Label speakers” at the top of a finished transcript and each paragraph gets a name you can change. It runs on your computer, like everything else.

• About 4 minutes per hour of audio on the test laptop, after the transcript is done.
• A 133 MB download the first time, then saved. The installed app saves it while it installs.
• Works on meetings recorded from one laptop or phone on the table, on phone calls, and on recordings up to 2.5 hours long. Names stay on the same people from start to finish.
• One person talking stays one speaker in 51 of 52 test recordings, apart from a short extra speaker now and then, usually marked “brief”. Merge it into the main speaker in two clicks.
• 2 to 11% of the speech got the wrong name in test meetings. Lines it is not sure about are marked, so you know which ones to check.
• Up to 8 people get their own name. With more, some share one.
• When people say their name in English, it suggests it (“My name is Eileen”). In 36 test meetings, 9 in 10 suggestions were right. Nothing changes until you click.
• Someone who speaks very little in a meeting can be labelled as another person. Move their lines to the right name.
• Now and then one person shows up as two speakers, for example someone sitting far from their laptop in an online call (1 in 13 test calls). Click either name to merge them.

Languages

• English, Spanish, Italian and Swedish. Choose the language before you transcribe.
• Spanish uses the more accurate model, the one Accuracy switches on. If Accuracy is off, it is a 240 MB download the first time you choose Spanish, then saved.
• Swedish uses its own model, KB-Whisper from the National Library of Sweden. It is a 240 MB download the first time you choose Swedish, then saved.
• Summaries are in English only, for now.

Privacy

• Runs on your computer. Nothing is uploaded.
• Summaries stay on your computer too. Key passages are exact quotes, with the time they were said. The written summary uses the AI built into Chrome and Edge. The first time, your browser may download that AI model from Google (Chrome) or Microsoft (Edge). Your recording and transcript are not part of that download.
• Every summary line links back to the part of the recording it came from. If nothing supports a line, it is marked no match.

What it runs on

• Chrome and Edge, on a laptop or desktop.
• About 2.3 GB of free memory with the standard model (3.9 GB with the graphics card). Recordings over an hour need about 10 MB more per minute.