Podcast production has exploded. The number of episodes published annually grew from roughly 3.9 million in 2015 to around 29 million in 2023.
Hours of valuable information are shared every day through this long-form audio. Yet despite how much useful information is buried inside podcasts, there still isn’t a comprehensive way to index them.
We have searchable indexes for news and databases for financial filings and academic papers. So what makes podcasts so much harder?
Having worked on indexing podcast content, I think the difficulty comes down to four problems: transcription, fragmented sources, content quality, and speaker identification.
Podcasts are audio first
Before you can make sense of a podcast, you first have to turn it into text. That sounds straightforward, especially given how much speech-to-text models have improved. But transcription is still far from a solved problem when you care about accuracy.
Proper nouns are a particularly difficult problem. Speakers constantly mention company names, product names, people, tickers, and industry-specific terminology that transcription models may not recognise. A company like Lyft, for example, can easily become “lift,” or Claude can become “cloud.”
This is a huge issue, especially when you are trying to build a searchable index. If the company name itself is transcribed incorrectly, the episode may never appear when someone searches for it. You need another layer that understands the context and corrects these transcription errors. The problem becomes even more challenging with different accents, speaking styles, recording environments, overlapping speakers, and poor audio quality.
The news on the other hand doesn’t have this problem. No form of processing is required as it’s already in the text form.
With podcasts, transcription is an additional computational step before indexing can even begin. At a large scale, that becomes a meaningful cost.
The source landscape is fragmented
The second problem is fragmentation. The barrier to creating a podcast is extremely low. Unlike traditional media, where a relatively small group of publications accounts for much of the trusted coverage, valuable podcast content can come from almost anywhere.
This is partly what makes podcasts interesting. You get perspectives, conversations, and expertise that would never appear in traditional media. But it also makes indexing them much harder. With news, there are a relatively limited number of publications that people consistently trust. You can get by with just indexing the major outlets.
Also Read: Asia’s AI trust gap: strong transparency, weak security and unclear data practices
Podcasts work very differently. A valuable piece of information might come from a huge show, a niche industry podcast, an independent expert, or a founder appearing on a tiny podcast with only a few thousand listeners. That means you cannot simply identify a few hundred trusted sources and call the job done. To build a useful podcast index, the coverage has to be dramatically wider. The long tail is not optional. It is often where the most interesting information lives.
That creates a scale problem that is easy to underestimate until you actually try to build it.
Content provenance and noise
The lack of editorial boundaries creates another problem: noise. AI-generated podcasts have become increasingly common. In our own work with financial podcasts, we have seen roughly 15 per cent of the content we encounter appear to be AI generated. Identifying and filtering this content is becoming surprisingly difficult.
I have worked around voice AI since 2023, and I used to think I had a good ear for identifying synthetic voices. I am far less confident today. As voice models improve, identifying AI voices has become a challenge.
There are other forms of duplication too. Podcasters frequently publish clips from other podcasts (reactions). An indexing system has to understand the difference. Is the person speaking actually a guest on this podcast? Is this a clip from another show? These problems can be solved using LLMs these days.
Not to forget, podcast advertising introduces complications. Podcasts increasingly use dynamically inserted ads, meaning the audio file itself can change depending on when or where it is played. Unlike an article sitting at a fixed URL with pretty much fixed content, podcast content is not always static.
Knowing what was said isn’t enough
The final problem may be the most important: speaker identification. Knowing that a statement was made is useful. Knowing who made it is far more useful.
Imagine someone saying that a particular company has an enormous competitive advantage. The meaning of that statement changes depending on whether the speaker is the company’s CEO, a competitor, an investor, a customer, or an independent industry expert.
The words might be identical but the perception will change accordingly. A useful podcast index needs to understand who is speaking and what their relationship is to the subject.
This is one area where modern AI models come in handy. Given enough context, they can often identify speakers, infer roles and resolve ambiguous references.
Also Read: Why Singapore’s AI finance race is now about data, not models
AI changes what is possible
A few years ago, building a comprehensive podcast index would have been theoretically unfeasible. You would need to transcribe millions of hours of audio, fix errors, identify speakers, and continuously process a huge stream of new episodes.
The only probable way would have been to hire troves of human annotators. LLMs and modern speech models change the economics of that problem. For the first time, it is becoming practical to turn podcasts from an audio format people have to manually consume into a structured data source that machines can understand. And that matters because the information inside podcasts is unusually valuable.
Executives explain how they think. Investors discuss their theses. Researchers describe work that may never appear in a paper. Founders talk about their companies in more detail than they would in a press release. Industry experts casually reveal insights buried inside hour-long conversations.
Until now, most of that information disappeared into the podcast feed after it was published. We are finally reaching the point where it doesn’t have to. Now that podcasts can be indexed, the more interesting question is what we can build once all of that information becomes searchable.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post Why podcasts are the next big data revolution appeared first on e27.