How to subtitle a video podcast
Long, conversational, two people talking over each other — the hard case, done in order.
A video podcast is the awkward case for automatic subtitling: it is long, it is conversational rather than scripted, there are at least two voices, and people interrupt each other. None of that is fatal, but the order you do things in makes a large difference to how long it takes.
Open the subtitle editor → Handles long files without a length limit. Expect a real wait on an hour-long episode.- 100% freeNo account, no minute quota, no paid tier.
- Nothing is uploadedYour video is read on your device and never reaches a server.
- No catch on the outputNo watermark, no length cap, no locked export formats.
- Updated in the openEvery fix is written down on the changelog page.
Transcribe the final cut, once
The single most expensive mistake is transcribing before the edit is locked. Every trim after that point shifts everything downstream, and re-timing an hour of subtitles costs more than the transcription did. Export the final video, then caption it.
Set the language explicitly rather than leaving it on auto. On a long recording, auto-detection samples the beginning, and an episode that opens with music or an ad read can start off on the wrong foot. Use the Accurate model if the audio is anything less than clean — for a conversation with crosstalk, the extra time is worth it.
Then leave it alone. An hour of audio on a laptop with WebGPU is a coffee break; without WebGPU it is longer. Lines appear as it works, so you can start reading the opening while the rest runs.
Speaker labels
The model transcribes what is said but does not separate who said it. Speaker separation is a different piece of machinery and is not implemented here, so labelling is manual — which sounds worse than it is, because in a two-person conversation you only need a label where the speaker changes, and you can hear those points as you review.
The convention is a short prefix at the start of the line: a name and a colon, or a dash for the second voice. There is no speaker field in SRT or VTT, so a prefix in the text is what everyone does. Do not label every line; label the changes, and only where the picture does not already make it obvious.
Crosstalk and filler
When two people talk at once, recognition picks whichever is louder and merges fragments of the other. There is no clean automatic fix. In practice you decide, per moment, whether the overlap carries meaning — often it is agreement noises that can be dropped entirely — and if it does, you split the line and time the two halves against the waveform.
The other judgement call is filler. Models transcribe "um" and "you know" fairly faithfully, and a subtitle track that keeps every one of them is exhausting to read. Broadcast practice is to remove filler and light verbal stumbles, because a reader processes text far faster than speech and the noise is more visible on the page than in the ear. Find and replace handles the repetitive ones in one pass.
Reading speed on fast conversation
Conversational speech runs faster than scripted narration, and the result is subtitle lines that are technically correct and physically unreadable. The reading-speed number beside each line is the thing to watch: when it goes orange, the line cannot be finished in the time it is on screen.
Two fixes, in order of preference: split the line so each half gets its own time, or trim filler to bring the character count down. Stretching the duration is the last resort because it pushes the line past the moment it belongs to. The auto-format button will handle wrapping and overlaps in bulk, then you deal with the handful of remaining flags by hand.
When it is done, export SRT for the video platform and plain text as well — the transcript is a separate, genuinely useful artefact for show notes and for search.
Frequently asked questions
Can it tell the speakers apart automatically?
No. Speaker diarisation is a separate model and is not implemented. You add labels manually at the points where the speaker changes.
How long does an hour-long episode take?
It depends entirely on your hardware. A machine with WebGPU is several times faster than the same machine on WebAssembly. There is no charge and no limit either way.
Should I remove ums and false starts?
Usually yes. Reading is faster than listening, so filler that passes unnoticed in audio becomes noise on screen. Broadcast practice removes it.
Is it safe to transcribe a recording I have not published yet?
The file never leaves your machine — it is read by the page and processed locally. Unreleased episodes and off-the-record material are exactly the case this suits.