Audio Transcriber
Speech in, text out, nothing uploaded.
or paste one from your clipboard
Off gives you one continuous block, which is easier to paste into something that will reflow it anyway.
This tool needs a 39 MB download, once
It is Whisper, OpenAI's speech recognition model, served from this domain rather than from Hugging Face. It is kept on your device afterwards, so this happens once and then the tool works offline forever. Transcription then runs several times faster than real time.
One time only. Nothing is uploaded — this is the model coming down to you.
Model ready on this device. Works offline, and nothing you load here leaves the browser.
How to use it
- Drop the recording in. The first run fetches Whisper, about 41 MB, once — after that this works with no internet at all.
- Press Transcribe. The model listens in thirty-second windows with a little overlap, so words are not cut in half at the joins.
- Read it, and fix what needs fixing. Names, jargon and acronyms are where every speech recogniser struggles, and it is faster to correct four words than to retype four hundred.
- Copy it to the clipboard or download it as a text file. If you want timings rather than prose, the subtitle generator produces the same transcript as SRT.
What the model is actually doing
Whisper does not recognise words the way older systems did, by matching sounds against a dictionary of pronunciations. It was trained on a very large quantity of audio paired with its transcript, and what it learned is closer to translation: given this stretch of sound, what text most probably accompanies it. That is why it punctuates, capitalises and spells out numbers without being told to — those were in the transcripts it learned from.
It also explains the model's characteristic strength and its characteristic failure. Because it is predicting probable text rather than decoding phonemes, it copes remarkably well with accents, mumbling and background noise, filling in from context the way a listener does. And because it is predicting probable text, when there is nothing to go on it will still predict something. Silence is the dangerous input, not noise.
Getting a better transcript
Clean the audio before transcribing, not after. Anything that makes speech easier for a person to follow makes it easier for the model too: the noise remover for a hissy or roomy recording, the silence remover to delete the passages where nobody is talking, and the volume booster if the recording is so quiet that the speech sits near the noise floor.
If the recording has music under the speech, running the vocal remover first sounds like a strange idea and works surprisingly well, because it is built to separate a voice from an arrangement.
Do not bother converting to a higher bitrate or a higher sample rate first. The model works at 16 kHz mono and the audio is resampled down to that regardless, so anything above it is discarded before the model sees it. A pristine 96 kHz file and a decent MP3 produce the same transcript.
What to check before trusting it
Proper nouns first. Names of people, places, companies and products are where errors concentrate, because the model is choosing the most probable spelling of a sound and an unusual name is by definition not probable.
Numbers second, particularly where they matter. "Fifteen" and "fifty" are a single vowel apart and both are plausible, so a figure in a transcript is worth confirming against the audio if anything depends on it.
And read any passage that looks unusually fluent for its position in the recording. A neat, well-formed sentence appearing where you remember a long pause is the signature of the model filling a gap rather than reporting one.