A podcast transcript pays for itself three ways: search engines can read it, deaf and hard-of-hearing listeners can use it, and you can search your own back catalogue. The transcription part is the easy step. The parts that go wrong are structural.
Two things that break on long audio
Chunk boundaries eat words
Speech models process audio in 30-second windows. Cut a two-hour episode into naive 30-second slices and every sentence crossing a boundary loses its middle. The output stays grammatical — two half-sentences join into a plausible whole — so nothing looks wrong.
This is not theoretical, and it is not limited to free tools. A hosted API tested on a 94-second file returned 191 words where the source had about 222: a 14% loss concentrated at the boundaries. Pinning different providers made no difference.
The fix is overlapping windows: process 30 seconds, step forward 25, reconcile the repeat. Good implementations do this by default. If you ever see words vanish at suspiciously regular intervals, that is what happened.
How to check: ordinary speech runs 120–150 words per minute. Multiply by your episode length. If the transcript is 15% short, you are losing content at boundaries.
Music and silence get filled in
Intros, stingers, ad breaks and long pauses contain no speech, but the model still produces its best guess at what belongs there. The classic artifacts are subtitle boilerplate absorbed from training data: “Thank you for watching,” a channel sign-off, a stray lyric.
How to avoid it: trim the musical intro before transcribing, or delete those lines afterwards. They cluster at the start and around ad breaks, so they are quick to find.
Speakers
Whisper produces one text stream and does not label who is talking. For a solo show this is irrelevant. For an interview show it is the main editing cost.
The practical options, in order of how well they work:
- Separate tracks. If you record each participant to their own file — as most remote recording tools do — transcribe each file independently. Each transcript contains exactly one person, and merging by timestamp gives you perfect attribution.
- Manual pass. For a two-person conversation, speaker changes are usually obvious from context.
- A separate diarization tool. More setup than most shows need.
If you already record separate tracks, option 1 is close to free and solves the problem completely.
What you can produce from one transcript
| Output | What it needs |
|---|---|
| Show notes | The transcript plus editing. Pull the three or four moments worth quoting. |
| Subtitles for video clips | SRT or VTT, both with timestamps. Most editors import either. |
| A searchable episode page | Plain text on the page. This is the SEO payoff — an hour of speech is thousands of indexable words. |
| Chapter markers | Timestamps from the transcript, grouped by topic. Manual, but the transcript makes it a five-minute job. |
For the episode page, resist pasting the raw transcript and calling it done. A wall of unedited speech ranks poorly and reads worse. Light editing — paragraph breaks, removing filler, a summary at the top — makes it useful to a human, which is also what makes it useful to a search engine.
Time and file size
An hour-long episode as MP3 is typically 30–60 MB, which browser-based transcription handles without an upload step. Processing runs at roughly a quarter to a half of real time on laptop CPU, so expect 15–30 minutes for an hour of audio. It runs in a background tab.
The first run also downloads the model once — 73 MB for English, cached afterwards. In one measured run, that download was 12.2 seconds of a 25.8-second total; from the second file onward it is gone.
Longer than about two hours, or working on a phone, and memory becomes the constraint before time does. Split the file at a natural break if you hit it.
Transcribe an episode — the audio is processed in your browser, so unreleased episodes stay unreleased.
Common questions
Should I publish the full transcript or a summary? Both, ideally: a summary at the top for humans, the full text below for search and accessibility. The full text is the part that makes an hour of speech indexable.
Does a transcript actually help podcast SEO? It is the only way search engines can read the episode’s content at all — audio is opaque to them. Whether it ranks depends on the same things any page depends on, but without it there is nothing to rank.
What about the ads? Transcribe the whole file and cut the ad reads afterwards. They are easy to spot and removing them makes the page better.
Can I do this for someone else’s podcast? For your own notes, research or quoting with attribution, that is ordinary use. Republishing a full transcript of someone else’s show as your own page is a different thing, and the transcript does not change the copyright position of the underlying work.