yescord

Transcribe a voice recording

Speech to text with timestamps, in English or 31 other languages, or translated into English, worked out on your own computer by a speech model. The audio never leaves this tab.

or drop a file anywhere on this page: a voice memo, WAV, MP3, M4A, OGG, FLAC, or a video, whose sound track is used.

The speech model and its runtime (54.7 MB) download from this site only when you press Transcribe, once per visit. Your audio is never uploaded.

How good is it?

This is a small model, and it makes mistakes. It is Whisper tiny.en, the smallest of OpenAI's speech models: 39 million parameters, stored here with 8-bit weights so it fits in a 54.7 MB download. Its model card reports a word error rate of 8.4% on clearly read audiobook speech (LibriSpeech test-clean) and 14.9% on harder audiobook speech (test-other). In practice that means:

  • Clear speech from one person close to the microphone (a voice memo, a dictated note, a read-aloud text) comes out mostly right: roughly one word in ten to twenty wrong, often small words or punctuation.
  • Names, places, brands and jargon come out wrong, often as a similar-sounding ordinary word: a surname becomes a noun, a drug name becomes nonsense. The model only writes words it saw often in training. Check every name and number.
  • Several people talking over each other, a phone call, a noisy room, music, or a strong accent make it much worse. It does not say who is speaking.
  • It sometimes invents text: in long silences or music it can repeat the last sentence or write a phrase nobody said. Skim the transcript against the recording.
  • Pick the language if you know it. English uses the English-only tiny.en. Any other language, “Detect the language” and “Translate to English” use the bigger multilingual base model (below), which is slower and a bigger download, and is much less accurate on Danish or Spanish than tiny.en is on English.

A worked example from this site's own test: a LibriVox volunteer reading of the Gettysburg Address (1 minute 57 seconds, public domain, a clear male voice), transcribed by this page in headless Chrome on a server on 2026-09-27, took 20 seconds and got 11 of Lincoln's 268 words wrong: a word error rate of 4.1%. That is a best case, one careful reader and a famous text. Almost every mistake was a sound-alike: “Four score” came out as “For score”, “hallow” as “halo”, “those who here gave” as “those who hear gave”, “living and dead” as “living in dead”, “nor long remember” as “no longer remember”. A bigger model would do better; the next size up (base.en) is about twice the download and twice as slow, which is why this page uses the tiny one.

Other languages, measured

Choose the language of the recording before you press Transcribe. Everything except English runs on Whisper base, the multilingual model one size up from tiny (74 million parameters, 8-bit): a 91.4 MB download the first time, against 54.7 MB for English. The table gives the share of words it got wrong (word error rate: wrong, missing and extra words divided by the words actually spoken) on one public clip per language, run through this page's own code in headless Chrome on 2026-09-29. One clip per language, one voice each: it tells you the order of magnitude, not a guarantee for your recording.

LanguageWords wrong, model this page usesWords wrong, the other model triedClip
English4.1% tiny.en (11 of 268)3.0% base (8 of 268)The Gettysburg Address read for LibriVox by Mark F. Smith, public domain, 117 s
Danish35.1% base (71 of 202)79.7% multilingual tiny (161 of 202)Ten read sentences from the NST Danish speech database (one woman from Funen, recorded 2000), CC0, 74 s: dataset alexandrainst/nst-da, test rows 1 to 10
Spanish23.0% base (26 of 113)41.6% multilingual tiny (47 of 113)The opening paragraph of the Spanish Wikipedia article Maracaibo, read by Wilfredor, CC0, 50 s (Wikimedia Commons, Maracaibo, Spanish Wikipedia Article Intro.wav)

Why base and not the multilingual tiny. Tiny multilingual is the same 55 MB as tiny.en, but on the Danish clip it got four words in five wrong, which is not a transcript. Base costs 36.6 MB more and about twice the time, and gets roughly two words in three right on the same clip. For English the difference between the models is small (4.1% against 3.0%), so English keeps the smaller, faster tiny.en.

What the mistakes look like. In Danish, base writes the sounds it hears as Danish-looking words: “Adskillige af de huse” became “Er det skælge af de huset”, “Hjørring” became “jørg”, “drejer til højre” became “drærer tilhøjere”. Place names and the odd word in the middle of a sentence suffer most; plain everyday sentences often come out whole. The Spanish clip lost its last few words and spelled names by ear (“Maracaybo”, “Neononbac” for Neu Nürnberg). Languages with more Whisper training audio (Spanish, German, French, Italian, Portuguese) generally come out better than Danish; languages near the bottom of the list, worse. None of the others in the menu has been measured here.

Detect the language runs the model once on the first 30 seconds and takes the language it rates most likely (the page shows which, and how sure it was); the whole recording is then transcribed as that language. A recording that switches language half way is transcribed in the first one.

Translate to English gives English text instead of a transcript, straight from the speech. It is rough: the Danish clip's “Der var både blodpølse og sylte på bordet” came out as “There were both bloodshed and silenced on the floor”. Good enough to tell what a recording is about, not to quote. It only translates into English, not between other languages.

How it works

  1. Decode at 16 kHz. The browser's own decoder (Web Audio, in an OfflineAudioContext at 16,000 samples a second, the rate Whisper was trained on) turns the file into samples. Stereo and multichannel files are mixed to mono.
  2. Load the model, only when asked. Pressing Transcribe the first time adds one script to the page, which loads transformers.js and ONNX Runtime Web and then the model's files, all from this site: 54.7 MB in total for English, of which the WebAssembly runtime is 11.1 MB and the two model files 40.9 MB; 91.4 MB for any other language, whose two model files are 76.9 MB. Nothing is fetched before you press it, and only the model the chosen language needs is fetched. The browser's ordinary cache may keep the files for a later visit; the page itself stores nothing.
  3. Listen in 30-second windows. Whisper reads exactly 30 seconds at a time. A longer recording is cut into windows that overlap by 5 seconds on each side (so a new window starts every 20 seconds), and the overlapping text is merged so words at the joins are not lost or doubled. The progress bar counts windows.
  4. Timestamps. Whisper writes a start and end time for each phrase itself, to the nearest 0.02 s. They are good to about half a second, enough to find a line in the recording or to time subtitles, but they are not word-by-word.

How long it takes. It runs on one processor core in WebAssembly. On a recent laptop a minute of speech usually takes somewhere between a few seconds and half a minute; a phone takes several times longer, and the first run also waits for the download. The page does not use the graphics card.

What you can save

  • TXT: the words alone, a new paragraph wherever the speaker paused for more than 1.5 seconds. Good for notes, a document, or pasting into an email.
  • SRT (SubRip): numbered subtitle cues like 00:01:04,200 --> 00:01:07,900. Almost every video editor and player (VLC, Premiere, DaVinci Resolve, YouTube upload) accepts it.
  • VTT (WebVTT): the web's subtitle format, the one an HTML <track> element plays. It uses a dot before the milliseconds instead of a comma, and starts with the line WEBVTT.

Subtitle lines are as long as Whisper's phrases, which can run to 20 words or so. If you need short two-line subtitles, split the longer cues in your editor.

What happens to your recording

Opening a file gives this tab a private blob: address for it. Decoding, recognition and the saved text files all happen in the tab's memory, and Save downloads them like any other download. This page has no upload code and no server to upload to, and it uses no browser storage, so closing the tab forgets the audio and the transcript. The only requests it makes are for its own files on this site (the page, then the model when you press Transcribe) and an anonymous, cookie-free page-view count (GoatCounter) that never sees your audio or your text. This site's automated test checks that a whole transcription run makes no request to any other address.

What it cannot do

  • No translation into any language but English, and no mixed-language transcript: one language per run.
  • No speaker labels ("Speaker 1:"), no word-by-word timing, no editing of the transcript on the page (open the TXT or SRT in any text editor).
  • No live captions while you speak: record first, then transcribe.
  • Memory: the page holds the audio as 16 kHz samples (about 4 MB per minute) plus the model (a few hundred MB while it runs). An hour-long recording works on a laptop; on a phone, keep it to a few minutes.
  • It opens only what this browser can decode: Apple Lossless only in Safari, old AMR phone memos in almost no browser. Record or convert to something else first; the trimmer can save any sound it opens as WAV.

The software and the model

All of it is stored on this site and served from here; nothing is loaded from Hugging Face, a CDN or anyone else. The versions are pinned and their licences sit next to them in the site's /vendor/ folder.

  • Whisper tiny.en, by OpenAI (weights released under the MIT licence; the Hugging Face model card lists Apache-2.0), converted to ONNX with 8-bit weights by Xenova (Xenova/whisper-tiny.en): an encoder of 10.1 MB and a decoder of 30.7 MB, plus its 2.1 MB tokenizer.
  • Whisper base (multilingual), by OpenAI (MIT weights; Apache-2.0 on the Hugging Face card), converted to ONNX with 8-bit weights by Xenova (Xenova/whisper-base): an encoder of 23.2 MB and a decoder of 53.7 MB, plus its 2.5 MB tokenizer. Used for every language but English.
  • transformers.js 3.8.1 (@huggingface/transformers, Apache-2.0): the audio features, the tokenizer and the 30-second window logic. Configured so it can only read models from this site.
  • ONNX Runtime Web 1.22.0-dev.20250409 (onnxruntime-web, MIT), the WebAssembly (CPU) build only: the part that runs the model.

What works where

Needed forChrome / EdgeFirefoxSafari (Mac)iOS SafariAndroid ChromeThis browser
Decoding and 16 kHz resampling (OfflineAudioContext)352514.114.535…
Running the model (WebAssembly in a module script)6160111161…
Finding the runtime (import maps)8910816.416.489…
Recording on the page (MediaRecorder)472514.11447…

Version numbers are the first release with the feature, from MDN browser-compat-data (the same source as the trimmer's table), checked 2026-09-27; they are not test results. The “This browser” column asks the browser you are using now.

What has actually been tested

One browser: headless Chrome on Linux, driven by test/transcribe.test.cjs in this site's repository. It drops the Gettysburg Address recording on the page, checks that no model file is requested before Transcribe is pressed, transcribes it, compares the words with the text, and parses the saved SRT and VTT files (every cue numbered, well formed, times in order). test/transcribe-languages.test.cjs does the same with the Danish clip: it picks Danish, checks the word error rate is no worse than the figure in the table above, then translates the clip and checks the result is English words, then tries “Detect the language” on the Spanish clip. Every request to another address is blocked and counted: there are none during the runs. Firefox, Safari, phones, noisy recordings and accents have not been tested by machine.

Made with AI tools. · Privacy