Today we’re releasing our initial support for the podcast:transcript tag. This is not generating transcripts but rather displaying the ones that are already in the feed, consistent with our goal of delivering the best UX possible while staying true to the creator’s content.
What we’re shipping
 |  |
We have full support for the four formats outlined in the podcast-namespace. So any standard JSON, SRT, or VTT file will "just work". See example file here corresponding to the screenshot above.
HTML support is a bit trickier but we've got it working well, even for podcasts with somewhat non-standard timestamps.
 |
 |
In addition, we are supporting XML transcripts like this one, and we're even rendering various plain text formats including this and this.
These are all being parsed on Castro’s backend and normalized for the client, offering a consistent user experience regardless of transcript origin. If you see a podcast that doesn’t seem to be rendering correctly, please email it to us. If you’re providing transcripts in a feed, using VTT or SRT is probably your best bet for the most consistent experience. If you’re not seeing a transcript in Castro that you expect, please email us and we’ll take a look.
Why aren’t we autogenerating?
I’m happy about what we’re releasing because it’s really important to me to show the podcast experience that the feed owner is trying to put out into the world. But there are drawbacks. For now, we can’t easily track the words to the audio since most transcripts do not include ads. Dynamic ad insertion in particular makes this difficult.
We could provide a more engaging UX here if we just generated our own transcripts. That’s the easiest way to solve the audio-word sync issue. Of course, there are cost concerns with that approach. In addition to cost, every layer of the stack adding their own transcripts doesn’t make a lot of sense to me.
- Podcast creators use their own tools for recording, editing, etc. Riverside, Zencastr, etc offer transcriptions.
- Podcast hosting companies have their own services and offerings. Again, transistor.fm and buzzsprout have transcription options.
- Operating systems can and indeed already do offer transcription of audio (Live Captions is very good on both Android and iOS)
- Do we also need the podcast app to generate another layer of transcripts on top of that?
Aside from this redundancy, transcripts generated on the client, even really, really good ones, cannot possibly know the creator’s intent. I see this issue repeatedly when I try offerings from other services. If a word is garbled or I can’t hear it correctly, that’s when I want to check the transcript. But in that situation, generated transcripts or live captions don’t have any better idea of what was said than I do and indeed they are often worse.
To me it’s far better to have the podcast provider tell you what they actually said. Of course, many podcasts are still going to use AI to generate their transcript, like Bloomberg and others are doing now, but at least they have the ability to correct mistakes on their end with this approach. This seems far superior to Castro compounding the problem by generating another layer of error-prone transcripts.
Future of the transcript tag
Adoption of this feature is not yet as widespread as I would like. Depending on how you slice the data, maybe 5-10% of new podcast episodes from popular feeds have transcripts. But that’s more than enough to be useful, and several large podcasts with extensive back catalogs have transcripts on every episode. However, there are two things that make me confident this number will grow and this is worth investing in:
- There are numerous options for transcript generation at every level now. It’s very low effort to tack on basic transcript support in your podcast recording software or from your hosting company.
- Industry adoption is widespread. Apple Podcasts shows the transcript tag when it is present. Pocket Casts, AntennaPod, and other apps also support this tag. Most major podcast hosting services support the tag.
Having access to the transcript is actually really handy. I was shocked how much I missed them once I spent a few days listening to shows with good transcripts. As more people experience this and more services make it easier to bolt transcripts on, I think we'll see transcript adoption continue to grow.
Making it better
This is only an initial release. We’re already working on improvements such as searching within the transcript, and we still want to pursue syncing audio to the text. We are exploring the best way to handle that, but I’m also open to ideas or discussion. If you are a podcast hosting company or creator who cares about transcripts and wants to help us provide a better experience to your listeners, please email support@castro.fm to discuss.