Free tool · Transcription · Runs in your browser
Audio to text
Short answer
Choose an audio file (MP3, WAV, M4A and more) and the tool transcribes it in your browser, with no upload. On English speech the Accurate model got 4.8% of words wrong in our test, about one in 21.
Nothing is uploaded: the audio is decoded and transcribed in your browser. The model downloads once and is kept by your browser for next time. Choose the language that is spoken; the model does not detect it.
Key takeaways
- Accuracy: 4.8% of words wrong on 48 minutes of read English with the Accurate model, 10.3% with the Fast one. The Whisper paper reports 5.0% and 7.6% for the same models.
- Timing: 99.9% of 885 subtitle lines landed inside the sentence they belong to, within half a second.
- Speed: 4.6 minutes of audio took 67 seconds on one processor core; an hour is about 14 minutes with the Accurate model and 8 with the Fast one.
- Silence is skipped: on silence the model invents words ("you"), so pieces quieter than -50 dBFS are not sent to it.
- Other languages are weaker: Spanish 13.7%, Portuguese 14.8%, German 21.1%, French 29.5%, Russian 30.4% with the Accurate model. Choose the spoken language; it is not detected.
- Private: your file is transcribed by your own browser and never uploaded. The model (80 MB) downloads once.
Measured: how accurate is it?
Download the text as TXT, or as SRT and VTT with timestamps.
Accuracy was measured with this page's own code and the same model files your browser downloads, run in Node.js on our server: 120 short recordings of read English (3 from each of the 40 speakers of LibriSpeech test-clean, 2,755 words) and 10 whole chapters joined from their sentences (48 minutes, 7,540 words). The word error rate counts wrong, missing and extra words against what was actually said, after lower-casing and removing punctuation; a number written as digits ("1830") where the reader said words ("eighteen thirty") counts as wrong, so the figures are a little harsher than the paper's.
| Model | Download | Short recordings | Whole chapters | Whisper paper |
|---|---|---|---|---|
| Fast (Whisper tiny) | 44 MB | 12.1% | 10.3% | 7.6% |
| Accurate (Whisper base) | 80 MB | 5.6% | 4.8% | 5.0% |
| Whisper small (not offered) | 252 MB | 4.0% | 3.8% | 3.4% |
The Accurate model got 4.8% of the words wrong on whole chapters and 5.6% on single sentences: about one word in 21. Longer recordings do better because every piece has full sentences around it. The Fast model is 44 MB instead of 80 MB but made 2.1 times as many errors on chapters. Whisper small would cut the errors to 3.8%, but it is a 252 MB download and took 1.8 times as long in our server run, so we do not offer it.
What to do with the transcript
Voice memos, interviews, lectures and meetings all work the same way: choose the file, choose the language, and edit the text before you download it. Timestamps per paragraph make a long interview easy to quote from.
If the audio is a podcast or an interview you want to promote, Koldflux can turn its strongest moments into Shorts with subtitles and carousels you review before export.
Long recordings: where each line lands
Whisper reads 30 seconds of audio at a time. The tool cuts your recording into pieces of 20 to 30 seconds, always at the quietest tenth of a second in that window, so a word is not cut in half, and shifts every line's time by the start of its piece.
To check the timing we built test recordings whose answer is known: each chapter is its sentences played one after another, so we know exactly when each sentence starts and ends. A subtitle line counts as placed right when its middle falls inside the sentences its words belong to, with half a second of slack. With the Accurate model 99.9% of 885 lines landed right (Fast: 99.9%).
When it gets it wrong, and what the tool does about it
Whisper has two known failure modes, and we hit both. On silence it invents a word: given 30 seconds of digital silence, the Accurate model wrote "you", and on quiet hiss it wrote "you". The tool measures the loudness of every piece first and skips pieces quieter than -50 dBFS; with that check both came back empty.
The second is looping: the model repeats a phrase until the piece ends. In our first run the Fast model turned one sentence of German into "ganze ganze ganze" and a Spanish one into "vamos, vamos, vamos". The Whisper paper's fix is to decode again with some randomness when the text compresses more than 2.4 times, a sign it repeats itself. The tool does a shortened form of that: if a piece compresses more than 2.4 times or has more than 6 words per second, it is read again at temperature 0.4, then 0.8. On the English chapters the Fast model needed a retry on 4 of 123 pieces and the Accurate model on none. If a piece still fails, the tool shows its time so you can check it.
What remains is ordinary mistakes: names, rare words and numbers. Whisper writes numbers as digits and spells names the way they sound, so read those back before you publish.
Other languages
Whisper is multilingual. We measured five languages on FLEURS, a set of sentences read by native speakers, 40 sentences each, with the same pipeline. Pick the spoken language in the tool: the model does not detect it, and reading Spanish as English produces a translation-like mess.
| Language | Accurate: words wrong | Fast: words wrong |
|---|---|---|
| Spanish | 13.7% | 30.9% |
| Portuguese | 14.8% | 37.4% |
| German | 21.1% | 49.9% |
| French | 29.5% | 58.5% |
| Russian | 30.4% | 56.9% |
Even the best of the five, Spanish at 13.7%, is clearly worse than English at 5.6%. Of Whisper's 680,000 training hours only 117,000 cover the 96 other languages, according to the paper, and the small models feel that most. For other languages use the Accurate model and expect to correct about one word in 7 or more. The other languages in the list were not measured here.
Speed in your browser
We timed the tool itself in headless Chromium on a 4.6 minute recording, on one core of a server processor with no graphics card; a recent laptop is usually faster. The model download happens once and your browser keeps it.
| Model | Download | Read the audio track | Transcribe | One hour of audio, estimated |
|---|---|---|---|---|
| Accurate (Whisper base) | 80 MB | 0.6 s | 66 s (4.2x real time) | about 14 minutes |
| Fast (Whisper tiny) | 44 MB | 0.6 s | 35 s (8.0x real time) | about 8 minutes |
Keep the tab open while it works. The hour estimate scales the measured time; the work is spread over pieces, so the progress bar moves steadily.
Audio files that work
We ran the tool on the same recording saved in 9 formats: MP4 (H.264 + AAC), MOV, WebM (VP9 + Opus), MKV, MP3, WAV, M4A (AAC), OGG (Vorbis), FLAC all worked. The tool reads the first audio track of the file and turns it into 16 kHz mono, the input Whisper expects; the file is read in place, not copied into memory, so long videos are fine.
How it works and why it is private
The tool runs OpenAI's Whisper speech recognition model (MIT license) in your browser through Transformers.js on ONNX Runtime Web. The model files are the 8-bit versions published by the onnx-community project and are served from this site; the first time you transcribe, your browser downloads 80 MB for the Accurate model or 44 MB for the Fast one and keeps them. Your recording is decoded and transcribed by your own device. It is not uploaded, so it can be a private call, an unreleased episode or a client video.
Whisper was trained on 680,000 hours of audio from the web, according to the paper, whose authors credit that scale for its robustness to accents and background noise. The two models here have 74 and 39 million parameters, the base and tiny sizes in OpenAI's table.
How the numbers were measured
Accuracy: Node.js on our server ran this page's own transcription code with the same ONNX files: 120 LibriSpeech test-clean recordings (CC BY 4.0), the first 10 speakers' first chapters joined with 0.3 seconds between sentences, and 40 FLEURS sentences (CC BY 4.0) in each of five languages. Word error rate is the edit distance between the output and the reference in words, divided by the number of reference words, after lower-casing and removing punctuation. Retries use randomness, so a rerun can differ by a few words. Speed and file formats: the tool itself in headless Chromium on a test build of this site. The results and scripts are kept with the page's evidence.
Questions people ask
How do I convert audio to text?
Choose the file above, pick the spoken language and the model, and wait for the progress bar. The transcript appears in a box you can edit; then download TXT, SRT or VTT.
Is my file uploaded?
No. The audio is decoded and transcribed in your browser. Only the model files are downloaded, once, from this site.
How accurate is it?
On read English the Accurate model got 4.8% of words wrong over 48 minutes in our test, and 5.6% on single sentences. Names, rare words and numbers are where most errors are.
Which languages does it support?
The list has 19 languages, and you must choose the one that is spoken. We measured five besides English: Spanish 13.7%, Portuguese 14.8%, German 21.1%, French 29.5%, Russian 30.4% of words wrong with the Accurate model.
How long does it take?
In our test, 4.6 minutes of audio took 67 seconds on one processor core, so an hour takes about 14 minutes with the Accurate model. A faster computer takes less.
Can it tell who is speaking?
No. It writes what is said, not who said it. Add speaker names yourself.
Can it transcribe a YouTube link?
No. It works on a file on your device. To transcribe your own video, use the original file you uploaded.
Why did it write something nobody said?
On silence or noise Whisper can invent a word, so the tool skips pieces quieter than -50 dBFS. On 30 seconds of piano music the Accurate model wrote "music". If a passage loops, the tool reads it again and marks it if it still fails.
What is the difference between the two models?
Accurate (Whisper base, 80 MB) made 4.8% errors in our test; Fast (Whisper tiny, 44 MB) made 10.3% and loops more often. Use Fast only for a quick look.
Related
Sources and how we checked
- Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (the Whisper paper), 2022 (read 2026-10-11)
- OpenAI Whisper: README (model sizes, MIT license) (read 2026-10-11)
- onnx-community/whisper-base on Hugging Face (the ONNX files this page serves) (read 2026-10-11)
- Transformers.js documentation (Hugging Face) (read 2026-10-11)
- LibriSpeech ASR corpus (OpenSLR 12), CC BY 4.0 (read 2026-10-11)
- Conneau et al., FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech, 2022 (read 2026-10-11)
Reviewed October 11, 2026. Accuracy was measured with the tool's own transcription code and model files on LibriSpeech and FLEURS; speed and file formats with the tool itself in headless Chromium. The runs and scripts are in data/seo/hub/evidence/2026-10-11/asr.