Transcribe Audio to Text
The first run downloads the speech model — about 45 MB for the fast tier, 81 MB for the accurate one, served by this site and cached by the browser afterwards. Nothing is uploaded, and the transcript is machine-made: proofread the wording and the timings before you publish it.
Options
Two tiers of the same model. The fast one is about 45 MB and transcribes roughly twice as quickly; the accurate one is about 81 MB and gets capitalisation and punctuation closer to right. Both run here, and the download happens once per browser. On a long recording, try the fast tier first to check the language is right.
The language being spoken in the recording. There is no automatic option: this build of the model does not detect language, and with none chosen it assumes English. A Chinese recording read as English does not fail — it produces fluent-looking nonsense, which is far harder to notice. Pick the language you actually hear.
This page turns a recording into plain text: paragraphs of what was said, with the timecodes left out. It is the same listening engine as the subtitle version of this page — same models, same thirty-second windows, same local-only rule — but the last step is different. Instead of writing a cue per sentence, it joins the sentences into paragraphs and drops the timing entirely, so what you get is something you can paste into a document, an email or a set of meeting notes.
Paragraphs break on the two things that actually separate thoughts in speech: a pause longer than a second and a half, or four sentences in a row. Only a gap-based rule would turn a five-minute monologue into one unreadable wall of text; only a rule based on sentence count would split a speaker's pause for breath into a new paragraph. Both together produce something that reads roughly the way the person spoke.
The work is local. The speech model is downloaded once from this site — about 45 MB for the fast tier, about 81 MB for the accurate one — and cached by the browser afterwards; the recording itself never leaves the tab. There is no upload step to skip and no account to create, which is what makes this usable on interviews that cannot be sent to a transcription service.
The output is a .txt file named after the source recording: film.mkv becomes film.transcript.txt. That distinct suffix is on purpose — a plain .txt sitting next to other text files should say what it is, and a transcript that was made by a machine is a different kind of document from meeting notes someone typed.
As with the subtitle version, what comes out needs proofreading, and the language has to be chosen by hand: this build of the model assumes English when told nothing, and getting that wrong produces fluent-looking nonsense rather than an error.
Frequently asked questions
How is this different from the page that exports a subtitle file as text?
The difference is where the timing comes from. The export page starts with a subtitle file that already has cues and joins their text into paragraphs — the timings existed and are simply being dropped. This page starts with audio that has no timing at all and builds it first, then drops it. That is why a recording with no subtitle track works here and not there, and why this page needs a model download while that one is instant.
Why are paragraphs split at a second and a half of silence?
Because that is roughly where speakers change thought. Shorter gaps fall inside sentences and would chop the text mid-clause; much longer thresholds join distinct topics into one paragraph. The four-sentence cap exists for the case where the recording has no silence at all — a rehearsed voice-over or a podcast recorded in one take — where pauses never come and a gap-only rule would produce a single block. The two rules together are the smallest set that handles both ends.
Can the transcript be used as evidence or as an official record?
Not as produced. It is a machine's best guess, and machine transcription garbles names, numbers and technical terms — precisely the words that matter most in a record. In a legal or medical setting the correct workflow is to use it as a draft that is then checked line by line against the audio, and to say so in the document. A cleaner transcript than the audio deserves is a liability, which is also why this page names the file .transcript.txt rather than something that looks like minutes.
What if the recording has music or several people talking over each other?
Overlapping speech is where any single-stream model degrades: two voices at once become one run-on sentence, and no speaker labels are produced. Background music is usually ignored, but loud music under speech can pull words out of the text or invent ones that were never said. If the result reads oddly in places, compare it against the audio at those points rather than trusting the flow of the text — a confident-looking sentence is not evidence that it was said.
Does the transcript keep the file's own structure, like chapters or track names?
No. Only the audio that is transcribed comes through; container metadata such as chapter markers, track titles and language tags is not part of the text. If you are transcribing one track out of a multi-track file, the page reads the audio from the file you dropped, and the paragraph breaks come from pauses, not from where the chapters were. Splitting a long recording into chapter files first is a reasonable workflow if you want those boundaries to show up.