Audio to Text Transcriber
Transcribe audio or video to text with Whisper, right in your browser: the file is never uploaded. Spanish, English and more, plus .srt and .vtt subtitles.
What this tool does
Upload an audio or video file —MP3, WAV, M4A, OGG, WebM or MP4— or record with your microphone, and Whisper, OpenAI’s speech recognition model, turns it into text. It works in English, Spanish, Portuguese and nearly a hundred other languages, and you can let it detect the language or pick it yourself.
You get the result in two shapes:
- Plain text, to copy or download as
.txt. - Timed subtitles, as
.srt(the format YouTube, Premiere, DaVinci Resolve and CapCut accept) and.vtt(the one web players use).
How to use it
- Drag the file in, click to choose it, or press Record with the microphone.
- Pick the audio language. If you know it, set it: automatic detection is almost always right, but fixing it avoids surprises with short or noisy clips.
- Pick the model: Fast (Whisper base) or Accurate (Whisper small). Each one shows how much it downloads.
- Press Transcribe. The first time the model is downloaded; while it works you see the text appear chunk by chunk.
- Copy the text or download the
.txt,.srtor.vtt.
Your audio never leaves your device
Typical transcription services upload your file to their servers. Not here: the audio is decoded and transcribed inside your browser. No account, no free minutes per month, no copy of your recording on another company’s disk. That matters when you are transcribing an interview, a consultation, an internal meeting or a voice note with someone’s details.
The only thing that travels over the network is the model, downloaded from Hugging Face the first time and kept in the browser cache. Next time it starts without downloading anything.
Requirements and download size
| Model | On CPU | With WebGPU |
|---|---|---|
| Fast (Whisper base) | ~80 MB | ~210 MB |
| Accurate (Whisper small) | ~250 MB | ~590 MB |
If your browser offers WebGPU (usually recent desktop Chrome or Edge), the model runs on the graphics card and is usually much faster. Otherwise it runs on the CPU: it works in any modern browser, just slower; on a phone, stick to the fast model and short clips. The tool tells you which one it will use and gives you a time estimate before starting.
The limit is practical, not fixed: the whole audio is loaded into memory and, on the CPU, anything over 10 minutes gets slow. For an hour of recording, use a desktop with WebGPU or split the audio. While it transcribes, keep the tab open and in the foreground: browsers throttle background tabs.
Where it gets it right and where it fails
Whisper is among the best open speech models, but the ones that fit in a browser are the small members of the family, so it helps to know where they stumble:
- Clear speech, one speaker, little noise: the text comes out almost clean.
- Background noise, music or several people at once: more mistakes, and sometimes skipped sentences.
- Names, brands and numbers: this is where it errs most. Always check them.
- No speaker separation: the text does not say who said what.
- Long silences: in stretches without speech it sometimes repeats a line or makes up a short one. If you see text nobody said, it usually sits in a silence.
The Accurate model usually makes fewer mistakes, especially with noise and uncommon vocabulary, in exchange for a download three times bigger and a longer run.
Subtitles for your videos
To subtitle a video, download the .srt and upload it to YouTube (under Subtitles) or import it into your editor. Each line carries its start and end time, so you only fix the text, not the sync. This tool does not burn subtitles into the video picture.
Frequently asked questions
Is my audio uploaded to a server?
No. The audio is decoded and transcribed inside your browser. The only thing that travels over the network is the Whisper model, downloaded from Hugging Face the first time and kept in the browser cache. You can check it in the network tab of the developer tools: no request carries your file.
How much is downloaded and how long does it take?
The fast model (Whisper base) is about 80 MB when it runs on the CPU and about 210 MB if your browser has WebGPU; the accurate one (Whisper small) is about 250 MB and 590 MB. It is downloaded once. After that, on the CPU, a powerful desktop transcribes with the fast model in about a third of the audio length and with the accurate one in about two thirds; an ordinary laptop can take two or three times longer. With WebGPU it is faster. The page shows an estimate before starting.
Is there a length limit?
There is no fixed limit, but there is a practical one: the whole audio is loaded into browser memory, and on the CPU anything over 10 minutes gets slow. For hour-long interviews or lectures, use a desktop with WebGPU or split the audio into pieces.
How accurate is it?
Whisper is among the best open speech models, but the ones that fit in a browser are the small members of the family. With clear speech and little noise the result usually needs few corrections; with background noise, overlapping speakers or heavy jargon it makes more mistakes. The accurate model reduces them at the cost of a bigger download. Always check names and numbers.
Can I transcribe a video?
Yes, if your browser can read its audio: MP4 and WebM work in Chrome, Edge and Firefox. The tool takes only the sound track. With the .srt or .vtt file you can add subtitles to the video on YouTube or in your editor; this tool does not burn them into the picture.
Does it tell speakers apart?
No. Whisper transcribes what it hears but does not separate speakers. The text comes out in timed segments, without labels for who said what.
Related tools
Others from the catalogue that pair well with this one.