BrowserTools
Home / AI / Audio to Text

Audio to Text

Turn a recording or video into text and subtitles, right in your browser, without uploading it anywhere.

Maintained by BrowserTools · Last updated 2026-09-29

Loading Audio to Text… If nothing happens, please enable JavaScript.

This tool listens to an audio or video file and writes down what is said. Drop in a voice memo, an interview, a lecture, a podcast episode, a meeting recording or a clip from your phone, and a few moments later you get the words as plain text, as a timestamped list, and as ready-made subtitle files in the two formats video players and editors expect: SRT and WebVTT. It works with the common formats (MP3, WAV, M4A, AAC, OGG, Opus, FLAC, WebM and MP4), and for less common containers such as MKV, AVI or MOV it pulls the audio track out first. You can queue several files and they are transcribed one after another, with a single .zip at the end holding the text and both subtitle files for each one.

Frequently asked questions

Is my audio uploaded to a server?
No. The speech model is downloaded to your browser and runs there, so the file is read and transcribed on your own device. Nothing is sent to this site or to anyone else, nothing is stored, and you can even disconnect from the internet once the model has loaded and it will keep working.
Why does the first run take longer?
The first time, your browser downloads the speech model, about 77 MB, from this site. It is saved in the browser's cache, so it only happens once per browser and later runs start right away. Clearing your browsing data removes it and it will simply be downloaded again next time.
How long does it take to transcribe a file?
It depends on whether your browser can use the graphics card. With a GPU (recent Chrome or Edge on most computers) an hour of audio can be done in a few minutes. On the processor alone, expect roughly real time or slower, so a 30 minute recording can take half an hour or more. The tool shows which one it is using, which part it is on and an estimate of the time left, and you can cancel at any point.
Which languages does it understand?
Whisper was trained on close to a hundred languages. Accuracy is best for widely spoken languages with lots of training data, such as English, Spanish, Portuguese, French, German, Italian and Dutch, and lower for languages with less. Automatic detection listens to the start of the file and picks one language for the whole recording; if a file mixes languages or the guess is wrong, choose the language yourself.
What does Translate to English do?
Instead of writing down the words in the original language, the model writes an English translation of what is said, with timestamps, so you can make English subtitles for a video in another language in one step. The translation is done by the same small model, so it gets the meaning across well but is not as polished as a dedicated translation tool.
Why did words appear that nobody said?
This is a known quirk of Whisper models: during long silences, background music, applause or very quiet speech they sometimes produce plausible phrases, or repeat the last sentence, because they always try to find words. The tool skips sections that are completely silent, but noisy or near silent passages can still trigger it. Trimming the silent or musical parts first, or checking those spots in the result, solves most cases.
What is the difference between SRT and VTT?
Both are plain text subtitle files with numbered or timed cues. SRT is the older, most widely supported format and uses a comma before the milliseconds (00:00:01,000). WebVTT is the web standard used by the HTML video element and streaming sites; it starts with a WEBVTT line and uses a dot (00:00:01.000). YouTube, VLC and most video editors accept either, so pick whichever your software asks for.
Can I transcribe a video file?
Yes. MP4 and WebM videos are read directly, and other containers such as MKV, AVI and MOV have their audio extracted in the browser first, which adds a one time download of the converter (about 30 MB). Only the sound is used, so the size of the video picture does not matter much, although very large files take longer to read into memory.
Does it tell different speakers apart?
No. Whisper writes down what is said but does not label who said it, so an interview comes out as one continuous text. Pauses of two seconds or more start a new paragraph in the plain text view, which often lines up with a change of speaker and makes it easier to add names by hand.

About Audio to Text

The transcription is done by Whisper, the open speech recognition model OpenAI released in 2022, in its multilingual base size. It understands close to a hundred languages, can work out which one is being spoken on its own, and can also translate speech in another language straight into English text. The model runs inside your browser: the first time you use the tool it downloads about 77 MB from this site, which your browser then keeps, so later visits start almost immediately. Your recordings never leave your device, which matters when the file is a doctor's appointment, a job interview, a family video or a confidential meeting. On a computer with a recent Chrome or Edge the work is done by the graphics card and is quick; otherwise the processor does it, which is fine for short clips but can take about as long as the recording itself, or longer, for long files.

Knowing how the model behaves helps you get better results. Clear speech close to the microphone works best, while heavy background music, several people talking over each other, strong accents and poor phone lines lower the accuracy. Choosing the spoken language yourself, instead of leaving it on automatic, avoids the odd case where a short or noisy file is guessed wrong. Whisper has a known habit of inventing plausible words during long silences, music or very quiet passages (a sentence like "Thanks for watching" appearing out of nowhere is typical), so the tool skips stretches that are completely silent, and it is worth reading the result before you publish it. The timestamps are accurate to within about a second, which is good enough for subtitles you can fine-tune in any subtitle editor.

Whisper learned from 680,000 hours of the internet

When OpenAI published Whisper in September 2022, the surprising part was not the model design, which was a fairly standard encoder and decoder, but the training data. Instead of carefully transcribed studio recordings, it was trained on about 680,000 hours of audio collected from the web together with the transcripts that already existed for it, roughly a third of it in languages other than English.

That messy, enormous dataset is why the model copes so well with accents, background noise and technical vocabulary without being tuned for any of them. It is also where its famous hallucinations come from: many online videos end with a line like "Thanks for watching, please subscribe" in their subtitles even when nobody says it, so the model learned that silence at the end of a clip is often followed by those words.

The model listens in fixed windows of 30 seconds, turning each one into an image-like map of frequencies over time before decoding it into text. Long recordings are cut into a series of these windows, and the timestamps it predicts inside each window are then shifted to their place in the whole file, which is exactly what this tool does in your browser.

Related tools

Audio Trimmer

Cut out the part of an audio file you want to keep, quickly and without losing quality.

Extract Audio from Video

Save the sound of a video as an MP3 or other audio file.

Image to Text (OCR)

Copy the text out of screenshots, scans and photos, right in your browser.

Background Remover

Cut out the subject of a photo and get a transparent PNG, right in your browser, without uploading it.
Support