Free Video Transcription

Correct, and it is the reason this tool exists in this form. The speech recognition model is downloaded to your browser and the audio is processed on your own machine. Nothing is transmitted, nothing is stored, and once the model is cached the page works with the network disconnected — which you can verify by turning it off. Almost every other free transcription service uploads your file to their servers, which for an interview, a medical consultation, a legal recording or an internal meeting is a meaningful thing to hand over.

Because the transcription happens on your machine rather than a server, and that means the speech recognition model has to be on your machine. There is no way around it: local processing requires local weights. The size depends on the model you pick and on your browser — the exact figure is shown next to each option — and ranges from about 43 MB to 395 MB. It happens once and is then cached. Services that need no download are doing the work on their own hardware, with your audio.

No — once, per model and per browser. The model panel says "Already downloaded" when the weights are on your machine, so you can see before starting whether a run needs the network at all. Two details worth knowing: the check is per browser, since the cache belongs to the browser rather than the computer, and it is per graphics path, because a browser using GPU acceleration needs the weights in a different format from one falling back to the CPU. Switching model size also means a separate download, which is why the panel prices each one.

Yes, and it is usually the right thing to do. Drag the handles under the player to select a span, and only that is transcribed. The panel shows what you have chosen in the same units the progress bar counts — "3 of 15 parts" — so the saving is visible before you commit rather than being a vague promise that trimming helps. Timestamps still refer to the original recording, so a passage selected from nine minutes in produces subtitles starting at nine minutes, not at zero, and an SRT will line up with the untrimmed video.

Good on clear speech and noticeably better with the larger models. Expect the usual failure modes of automatic transcription: proper nouns, technical jargon, homophones, heavy accents, crosstalk and background noise. It is well suited to subtitles, notes, searchable archives and first drafts. It is not a substitute for a human transcriptionist where accuracy is legally or medically material, and no automatic system currently is.

There is a real progress bar with a time remaining, not a spinner. The audio is processed in 30-second windows, the bar counts the ones finished against the total, and the estimate is the median time of the windows already done — so it settles after the first couple and does not lurch when one window happens to be slow. Speed depends heavily on your browser: with GPU acceleration a recording transcribes several times faster than its own length, while the CPU fallback is closer to real time or slower. The page says which one it is using before you start, and the work runs on a background thread so the page stays usable throughout.

Balanced for most things; it is the default for good reason. Fast is worth it on a metered connection or for clean, clearly-spoken audio like a solo screen recording. Accurate is worth the larger download and the extra time for difficult material — strong accents, several speakers, background noise, or recordings where names and terminology matter. All three are the same Whisper architecture at different sizes.

Yes, in both standard formats. SRT is the most widely supported and is what most video players and platforms expect. WebVTT is the web standard, used by the HTML video element and required by some platforms. Both carry the timings from the transcript, and lines are wrapped to around 42 characters over at most two lines, which is the conventional limit that keeps text inside the safe area rather than overflowing the picture.

Treat them as a starting point that is quick to correct, not as a finding. Voices are grouped by acoustic similarity — the tool measures the character of each segment and clusters segments that sound alike. That works well when voices differ clearly, such as two people alternating in a clean interview. It degrades on similar voices, very short segments and background noise, and it cannot represent overlapping speech at all, so a segment where two people talk at once gets a single label. Telling it how many people are speaking is markedly more reliable than letting it work that out. Every label can be reassigned afterwards.

Yes, into English. Whisper can transcribe non-English speech directly into English text in the same pass, so it costs no extra time. One caveat: you get the translation instead of the original, not both — run it twice if you need each. Translation into languages other than English is not something the model does.

The practical limit is three hours, which is set by memory rather than policy: decoding holds the audio track uncompressed, so a very long recording can exhaust the tab before transcription even begins. There is no limit on file size, since the file is never uploaded — a 4 GB video is fine as long as the recording itself is within that duration.

Anything your browser can decode the audio of: MP4 and MOV with AAC, WebM, MP3, M4A, WAV, OGG and FLAC. Video files are fine — only the audio track is used. If a file is rejected it is almost always an unusual audio codec, and converting it to MP3 or WAV first will fix it.

Yes, directly on the page. Click any line to correct it and the change flows through to every export format. This is deliberate: every automatic transcript needs a pass for names and jargon, and a tool that hands you an uneditable block of text just forces that work into another application. Clicking a timestamp jumps the video to that moment so you can check what was actually said.

Once the model is cached, yes, completely. The first run needs the network to fetch the model; after that the page and the model are both cached and the whole tool runs disconnected. That is a genuine property of doing the work locally rather than a claim about a privacy policy.