Subtitles for Video and Video-to-Text Transcription

Upload the file and press “Generate captions”. When recognition finishes, a “Copy” button and a “TXT” download appear above the transcript — you can take the text as flowing prose or with timecodes. The “Plain text” toggle shows the result as an ordinary document, ready to paste into notes or an article.

SRT for almost everything: YouTube, players, editing suites all read it, and it carries plain timed lines. VTT is the web format — it is what a browser expects in a video track element. ASS is the one to pick when the styling has to survive the trip: fonts, positions and colors live inside the file, so a caption designed here still looks designed when someone else opens it.

It listens first and decides: Whisper detects the language from the audio itself across more than ninety, so a clip you did not record yourself still comes out in the right one. Russian goes through a model trained for it rather than the general one. If the recording switches languages, or a heavy accent throws the detection off, set the language by hand before the run and the guesswork disappears.

Yes. Silero VAD marks the stretches of real speech before recognition, so silence in intros, pauses and outros is skipped and accuracy holds up on long files. Text appears in the editor as decoding progresses — you do not have to wait for the end to start reading and fixing. Videos over 30 minutes are more comfortable on a desktop with 8 GB of RAM.

Yes — “Download MP4 with subtitles” builds a file with the captions drawn over the picture, so they show in any social feed and in players that cannot read SRT. The build runs in the browser through FFmpeg WASM: video is encoded to H.264 and the audio stream is copied without re-encoding.

Click the timecode to the left of any line in the transcript and type the new time, or nudge it with the arrow keys (Shift moves in one-second steps). The video jumps there as you edit, so you can hear the result immediately. Word timings are re-derived from the line, so the karaoke highlight stays in step.

Whisper derives word boundaries from cross-attention decoding, with roughly ±200 ms accuracy — enough for karaoke highlighting and for ordinary subtitles. If decoding is slow, turn on fast mode and words are spread evenly inside each segment. A line start time can also be nudged by hand in the transcript.

VTT export is the standard closed-caption format for web players (HTML5 track, YouTube, Vimeo) accepted by WCAG, ADA and EAA checks. Captions carry timecodes, speaker labels when the detector is on, and clean punctuation. The burned MP4 adds outline and shadow for contrast and keeps captions inside platform safe zones.

Open “Glossary” in the recognition settings and list your terms, one per line. They are passed to the model as an initial prompt, so product names, surnames, acronyms and professional jargon come out spelled correctly. This is the single biggest accuracy gain on technical talks and branded content.