I timed every free in-browser Whisper transcriber

Mastershot's audio to text runs OpenAI's Whisper in your browser, so your recording never leaves your computer. There are a bunch of other free sites that do the same thing. Since I'm running it in the browser, I have to use a smaller Whisper model so I selected Whisper Base (multilingual) model. For the purposes of this test all the other websites that I tested, I used the same exact model. The reason I wanted to compare to other websites was because I heavily optimized the processing architecture for longer audio recordings for accuracy and speed.

TL;DR: On an 8-minute clip, the three fastest tools finished within 3 seconds of each other, and almost every tool got the same words right. On a 57-minute recording, Mastershot finished in 2 minutes 3 seconds, and the next two took 3 minutes 21 seconds and 3 minutes 26 seconds respectively. Additionally the other two sites had some repeated sentences and skipped ones as well. Mastershot had zero repetitions.

Below are all the numbers and the methodology I used to test it all.

The rules

• Free, with no limits. If you'd have to pay to transcribe an hour of audio, the tool isn't in the test since the paid ones generally run on a server on a much larger model. The ones I left out for that reason are listed further down, with their limits.
• The same model. Every tool was set to Whisper base, the multilingual one, which is what Mastershot runs.
• In the browser. Everything here transcribes on your own computer. Nothing is uploaded.

The recording

I used a LibriVox recording of The Art of War, read by Moira Fogarty in Lionel Giles's translation. It's in the public domain, so anyone can download it and run the same test, and the book's text gives me the exact text I need to compare with. I scored against Project Gutenberg's edition without commentary, plus the LibriVox announcements she reads at the start and end of each part ("This is a librivox recording, etc.).

I ran two tests:

• 8 minutes 25 seconds: the introduction and the first two parts, 1,270 words.
• 57 minutes 10 seconds: the first eleven parts, 8,804 words. Since there were some results in the 8 minute run that took a long time I only ran the top 3 on the longer recording.

How I timed it

Everything ran on a MacBook Pro with an Apple M2 Pro and 32 GB of memory, plugged in, in Chrome 155 on macOS 26.6. I ran one tool at a time with nothing else running, waited 20 seconds between runs, and shuffled the order each round.

Each tool had already downloaded its model before the timed runs, so the times don't include the download. The clock started when I clicked the tool's start button (on Zalt, which starts as soon as you pick a file, when the file was picked) and stopped when the full transcript was on the page. Each tool ran three times and I report the middle time.

How I scored it

I counted the word error rate (WER): the words a tool got wrong, left out, or added, as a share of the words in the known text. Lower is better. Before comparing, both texts go through Whisper's own English normalizer, which ignores capitals, punctuation, and how numbers are written, so only the words count.

Two things about the known text.

1. The reader skips three sentences that are in the book. Every tool leaves them out too, and the audio has no pause where they'd be, so I took them out of the key.
2. In a handful of places every tool hears a slightly different word from the book ("mantles" for "mantlets", "income" for "incomes"). I kept the book's word which caused a slight increase in WER.

The 8-minute clip

Most of the tools heard the same words, which makes sense, since it's the same model. The differences come from what each tool does around it:

• whisperweb.app heard as well as the others, but its TXT download joins each line to the next with no space between them ("a matter of life anddeath", "can onno account be neglected"). If I consider that as accurate, it scores 3.9%. It's built on Xenova's open-source Whisper Web.
• Zalt left out about 150 words in a few long stretches, including most of the first paragraph.
• Knackpad was the most accurate, but it runs Whisper on the page itself in the main thread, so the tab freezes until it's done. It took three times as long as the fastest tools.
• whisper.cpp's demo is the official browser build of whisper.cpp. Its site lets it use 8 processor threads. It's accurate, but it needed more than twice as long as the fastest tools.

The three fastest tools all run Whisper on the computer's graphics chip through WebGPU. On a recording this short, they're about even.

The 57-minute recording

Over an hour, the gap opens. whisperweb.net and EarScribe both work through the recording in 30-second pieces that overlap a bit. In cases where two pieces meet, sometimes a sentence is written twice (e.g."for every victory gained you will also suffer a defeat" comes out twice in both) and sometimes not at all. Both also leave out the last sentence, which the recording cuts off partway through.

EarScribe at 12:43 of the 57-minute recording. The line ends with a second copy of "for every victory gained you will also suffer a defeat."

Mastershot is faster on long recordings because on a computer it splits them at pauses in the speech so there's no overlap, into max three parts, and transcribes the parts in parallel with workers. (On phones, iPads and in Safari it stays as one worker) It still had no duplicated parts. It did miss two short stretches: "So much for tactical dispositions" at the end of the fourth part, and "Recorded in Toronto, Ontario, by Moira Fogarty, October 2006" after the eighth.

The same passage in Mastershot, written once.

What you get if you don't change the settings

I set every tool to Whisper base for a fair test, but that isn't how most of them start. whisperweb.app, whisperweb.net, Zalt and Knackpad all start on Whisper tiny, a smaller and less accurate model. EarScribe starts on base. Mastershot only runs base.

Tools I left out because they aren't free

I checked each of these on its own pricing page on October 14, 2026. All of them let you try transcription for free, but you would actually need to pay when it comes to a 57 minute recording.

Monthly prices are for paying a year at a time, where that's cheaper. SoundTools is free with no limits, but it only offers Whisper small and an English-only distilled model, not base, so it wasn't a fair match for this test.

Check it yourself

Everything I used is here, so you can run the same test on your own computer, or on a tool I missed:

• The 8-minute clip (6 MB)
• The answer key for the clip and for the 57 minutes
• The scoring script, which needs Python with jiwer and whisper-normalizer
• Every tool's transcripts, as each one gave them to me

The 57-minute recording is the first eleven parts of the LibriVox recording that are all joined together and cut off at 57 minutes 10 seconds, partway through part eleven. If you find a free tool that runs Whisper base in the browser and I left it out, tell me through the Feedback link at the bottom of the page and I'll test it.