A conference talk, a long interview, a recorded lecture, a webinar you sat through nine months ago — there is a great deal of useful material that exists only as video. It is hard to search, hard to quote, and effectively impossible to use as a research source without watching the whole thing again with a notepad open.
Podcast Creator Studio can take a video link, turn the speech into text, and file that text as a source your episodes are written from. The transcript is also yours to download as a plain .txt file, whether or not you ever generate an episode from it.
Where it lives
Open any podcast and go to Sources → Uploaded materials. Above the file picker there is a Transcribe audio/video panel. It accepts two things:
- A link. YouTube and Facebook are the common cases, but the same downloader handles Vimeo and roughly eighteen hundred other sites.
- A file. Any audio or video you already have — MP3, M4A, WAV, MP4, MOV, MKV and the rest, up to 500 MB.
Transcription runs in the background. A long video takes minutes rather than seconds, so you can leave the tab, and the panel reports progress as it goes.
Captions first, because they are free and often better
Before transcribing anything, the studio checks whether the video already carries subtitles. Most YouTube videos do. Fetching them takes a few seconds, costs nothing, needs no API key at all, and when the subtitles were written by a human they are usually better than what speech-to-text would produce — correct spelling of names, real punctuation, no guessing at technical terms.
Human-written subtitles are preferred over auto-generated ones where both exist. Auto-generated captions repeat each line as the next one scrolls into view, so those duplicates are stripped and the text is reflowed into sentences before it is stored — otherwise a transcript comes out roughly twice as long as the speech and reads like a stutter.
When there are no captions
Plenty of video has no subtitle track at all — most Facebook video, older uploads, anything you recorded yourself. In that case the audio is downloaded, normalised to 16 kHz mono, and run through a speech-to-text engine.
The default engine is local Whisper, running on the server itself. It is worth being precise about what that means:
- No API key and no per-minute cost. Nothing to sign up for, nothing metered.
- The audio never leaves the machine. For licensed material, unreleased recordings, or anything under an NDA, this is the difference between a usable workflow and one legal will not approve.
- It is fast enough. On a well-specified server a nineteen-minute video transcribes in a couple of minutes — comfortably faster than real time.
If you would rather trade money for speed, hosted engines are available too: Groq, Deepgram, ElevenLabs Scribe and OpenAI Whisper all work with your own API key. You choose a primary and a fallback under Settings → General → Transcription. The fallback only runs if the primary fails, so nothing is charged unless it is actually used.
The transcript becomes a real source, not a note
This is the part that matters. The finished text is not dropped into a scratch field somewhere — it is filed through the ordinary upload path, exactly as though you had uploaded a PDF:
- It appears in the Uploaded materials table with a word count and status.
- It is split into overlapping chunks and indexed into that podcast's retrieval store.
- At generation time, the passages relevant to each segment are retrieved and the script is written around them — and fact-checked against them.
- It is downloadable as a
.txtfile whenever you want it.
In other words, nothing downstream treats a transcript differently from any other document. A podcast whose sources are two PDFs, a scraped site and a transcribed conference talk simply has four sources.
A note on what to transcribe
This is a research tool, and it is worth saying plainly: transcribing someone else's video does not give you the right to republish their words as your own episode. What it is genuinely good for is the case where you need the substance:
- Your own recordings — talks you gave, interviews you conducted, lectures you own.
- Material you have permission to use, or that is openly licensed.
- Research and reference, where the episode discusses and cites a source rather than reproducing it.
Because the script is grounded in retrieval and then fact-checked, the natural output is an episode that discusses what a source said — which is both the more useful outcome and the more defensible one.
A worked example
Say you run a weekly show about a technical field, and a major conference just published thirty talks. Transcribe the six that matter. Each one lands as a source. Your next episode is written from the actual content of those talks — the specific claims, the numbers, the disagreements between speakers — rather than from a model's general impression of the field. The hosts can reference what was actually said, because it is in the index.
That is a different kind of episode from one written off a topic prompt, and the difference is audible.