Podcast transcription

Nearly every podcast puts music under the host at some point, so in August 2026 we tested how much that costs. We mixed an instrumental bed under the same spoken clip at four levels, from quiet behind the voice up to as loud as the voice itself, and ran every version through the same transcription request.

Every word survived at all four levels. What changed was punctuation: as the bed got louder, the engine placed commas where it had used full stops, because the pauses it listens for to end a sentence were partly masked. No words were dropped or misheard, even in the version where the music matched the voice in volume.

The honest limit on that test matters more than the result. Our bed was a synthesised instrumental tone, which is a fair stand-in for the simple loops used under podcast intros and outros, and a poor one for a dense track with drums and vocals. A song with a singer competing with your host is a genuinely harder problem, and we did not test it. Treat this as evidence that a typical intro bed is not the thing to worry about, not as a claim that music never matters.

The practical version: if you have the pre-mix export from your editor, the one with the voice track alone, upload that instead. It removes the question entirely and takes no longer.

A raw transcript is a poor web page. It repeats, it wanders, and publishing it unedited is the fastest way to make an episode look worse in writing than it sounded. What works is treating the transcript as source material: pull the three or four claims worth quoting, keep the timestamps so listeners can jump to them, and write the connective tissue yourself.

The AI layer helps with the first pass on paid plans, drafting summaries and key takeaways from your actual transcript rather than from a guess about what the episode covered. It is a starting point that still needs your judgment about what mattered, which is the part no tool can do for you. There is a fuller walkthrough in our guide on turning a podcast into a blog post.

Guest names are the one thing worth fixing before anything else. They appear throughout an episode, they are exactly the words speech recognition finds hardest, and they are the most embarrassing thing to get wrong in public. Add them to the vocabulary box before uploading and most arrive correct; fix any stragglers once in the editor and the correction carries into every export.