How do you transcribe a podcast?

To transcribe a podcast, run the audio through an automatic tool, check every line against the recording, and add a name for each speaker. Then publish the checked file in a format your host accepts. The first draft takes minutes. The check and the speaker names take the most care, because listeners will quote the episode and a misquote becomes a correction.

Apple makes a transcript, but it does not name the speakers

If you publish on Apple Podcasts, Apple generates a transcript for you. Apple's support page says that it "automatically generates transcripts after a new episode is published" ¹, and the app labels the result as automatically generated. The page also says that a VTT file lets you identify every speaker on each line ². The speaker names come from that file, which you supply.

Apple's list of languages for its transcript display is also limited. It names English, Danish, Dutch, Finnish, French, German, Italian, Norwegian, Portuguese, Spanish and Swedish for iOS 17.4 or later ³. If your show is in another language, check the current list before you assume the transcript will appear.

Make your own checked file to get speaker names

Provide your own transcript when you want names, quotes you can trust, or a language outside Apple's list. The steps are:

1. Export a clean audio file. Use the master recording, not a compressed copy, and keep one track per speaker if your recorder can save them.
2. Run an automatic tool. Any tool that gives timestamps will do, but test it on a few minutes of your own audio first.
3. Check every line against the audio. Play at normal speed, and fix names, product names, numbers and negatives first.
4. Add a name to every speaker turn, using the same name throughout the episode.
5. Save as a WebVTT file. The glossary entry on WebVTT explains the format, and the entry on timestamps explains the time codes.
6. Upload or link the file through your host, and confirm the transcript displays on the episode before you announce it.

Some recording platforms also offer built-in transcripts. Check their help center for the plan, language and speaker settings before you rely on one, because those settings can change between plans and releases.

Multi-speaker conversation is where automatic errors cluster

A podcast with two or three people talking over each other is harder than a read speech. On the CallHome test set, which records conversations between friends and family, Whisper Large V2 had a word error rate of 17.6 percent ⁴. The glossary entry on word error rate explains how that figure is calculated. A conversation transcript needs the most checking exactly where people interrupt each other.

Automatic speaker labels are a separate problem. OpenAI's speech-to-text guide says its speaker-identification model should be used only when you need to identify who speaks during different parts of a recording ⁵. Do not assume a tool has correctly assigned each line to the right guest. Listen to each speaker change while you check.

Wollin-Giering and colleagues recommend reviewing the text of any transcript to ensure accuracy, whatever tool produced it ⁶. For a podcast, the review is the step that protects your guests and your credibility.

It depends: what the show needs

Keep the checked text, the audio and a note of which tool made the draft. If an episode is corrected later, you need the audio to show what was said.

What to do, in order

1. Record one track per person where you can, and keep the master file.
2. Generate a draft with a tool you have tested on your own voices.
3. Check every line against the audio, and name every speaker.
4. Save as WebVTT and upload it through your host.
5. Confirm the display on the published episode before you share it.

For interviews where the recording is part of a research project, read our question on using AI transcription for research interviews, which covers consent and data protection. For the broader method, see our guide to interview transcription, and our guide to timestamps and speaker labels for the label formats.