Transcribe audio to text, free and private.

At a glance

Unlimited transcription, because nothing is uploaded

The file is decoded in this tab, folded down to one channel and resampled to 16 kHz, which is the only rate Whisper understands. That happens with the same windowed-sinc resampler the rest of AudioSaw uses rather than the browser's built-in conversion, so there is no aliasing for the model to mistake for sibilance. The audio is then handed to a background worker that runs the network in 30-second windows with five seconds of overlap on each side, and stitches the windows back together where the overlaps agree.

None of that leaves your machine. Most free transcription sites upload the recording and run Whisper on their own server, which is why they want an account and cap you at a few minutes: someone is paying for the GPU. Here your GPU, or failing that your CPU, does the work, so there is no account, no minute cap and no copy of your audio sitting on a stranger's disk. That matters most for exactly the recordings people most want transcribed — interviews with sources, therapy and medical dictation, legal calls, lectures you were not supposed to record, a draft of a song.

Which Whisper model to pick

Base is the default because it is the best trade for most recordings: clear speech in English and the major European languages comes back with very few mistakes, and the download is a one-off. Tiny is the one to use on a phone or an older laptop, or when you only need to search a recording rather than quote from it; it is several times faster and noticeably sloppier with names, numbers and accents. Small is the most accurate option here and the one to choose for a heavy accent, a noisy room, or any language other than the big European ones — Hindi, Persian, Tamil and Vietnamese all improve markedly from base to small. It is a large download and slow on a CPU.

The page probes your browser before you start and says whether it can use the GPU. With WebGPU (current Chrome and Edge on a desktop) Whisper runs several times faster than on the CPU; the words are the same either way. Without it, the CPU path splits the work across your cores, which is why the tab needs to stay open and why a laptop fan may spin up.

Video to text: SRT or VTT subtitles

Both files hold the same cues — a start time, an end time and a line or two of text. SRT is the one nearly every editor and platform imports: YouTube Studio, Premiere Pro, DaVinci Resolve, Final Cut (via its caption import), CapCut and VLC all take it. VTT is the web's own format and what an HTML5 <track> element needs. Lines are wrapped at 42 characters, the broadcast convention, and every cue is kept on screen for at least a third of a second so nothing flashes past unread.

Whisper times phrases, not words, so a cue covers a natural chunk of speech of a few seconds. That is the right granularity for subtitles. It is not the right tool for karaoke-style word highlighting, and the page does not pretend otherwise.

What it gets wrong

Proper nouns are the commonest error: a product, a surname or a place it has not heard spelled is written the way it sounds. That is why the transcript is an editable box — fix the names before you download the text. Subtitle files keep the original wording so their timing stays aligned.

Transcribe an interview with speaker labels

Whisper itself does not tell speakers apart; tick "label who is speaking" and a second step does. It is pyannote's method, run in your browser: a small segmentation model (pyannote segmentation-3.0) marks where speech is and when the voice changes, a speaker-embedding model (WeSpeaker ResNet34, CC-BY-4.0) turns each stretch of speech into a voiceprint, and the voiceprints are grouped into people. Each line of the transcript then gets "Speaker 1:", "Speaker 2:" and so on, and you can type real names in once. Subtitles get the name at each change of speaker. On test conversations it labelled over 99% of the speech correctly with two people and 92% with three, and counted them right when left to count; tell it the number of people if you know it. Short interjections ("yes", "right") are the usual mistakes, and two people with very similar voices may be merged. Crosstalk, where two people talk at once, tends to keep one voice and lose the other.

What else it gets wrong

On long stretches of music or silence it can invent words — usually a repeated phrase or a polite "Thank you." It was trained largely on subtitled video, and that is what subtitles say when nobody is talking. Trimming a long music intro with the cutter, or tightening the gaps with auto-cut silence, removes most of it. Background hiss and hum hurt accuracy more than people expect; a pass through noise reduction first is often the single biggest improvement.

"Translate into English" is Whisper's own translation, done in the same pass. It is good enough to understand a foreign-language recording and to make rough English subtitles; it is not a translator you would publish from unchecked, and it only goes into English.

Compared with the alternatives

Otter, Rev, Descript and the transcription built into meeting apps are convenient and label speakers too, but they all work by uploading your audio and most charge by the minute beyond a free allowance. Running Whisper yourself in Python, or with a desktop app such as MacWhisper, gives you the larger models and is the better choice for hours of material every week. This page sits between the two: the same model family as the desktop route, with nothing to install, and the privacy of doing it on your own machine. If you do want to run Whisper elsewhere, audio for Whisper produces the 16 kHz mono WAV it is fed internally.