South African talk radio is now searchable

South African talk radio is the last big pool of news that nobody can search. Locl is an attempt to change that: a platform that listens to the live streams of the country’s major English-language talk stations, transcribes everything, indexes it and lets you look up any name, brand or subject and find every time it was mentioned on air.

The stations it ingests are 702, Cape Talk, Power FM and SAfm, with plans to include non-English platforms in time.

“I love all things news, and I’ve always loved all things news,” its creator, Shaun Trennery, told TechCentral on Friday. “With the advent of AI, you’re able to get pretty much any information. The only thing that was missing was radio.”

What interested Trennery was not the bulletins. It was the callers, the interviews and the arguments – the part of radio that has never been indexed, and which disappears the moment it is broadcast. Podcast feeds give you whole shows with no way in. “There aren’t any highlights or anything like that.”

Trennery has been writing software for about 25 years. But he built locl without writing any code himself. “It’s 100% Claude Code,” he said. “We’re at the stage now, we’re not even looking at the code anymore.”

The platform took three to four months, built in the gaps of a working day. “While I’m working throughout the day I have an idea, or on my run I have an idea, and then I just give it to AI, and it iterates and it creates a PR (pull request) that I’m able to review and merge in.”

A listening tool runs per station, capturing the live stream in half-minute chunks. Each chunk goes through Whisper, the open-source speech recognition model, and the text lands in what Trennery calls a massive index. On top of that sits a retrieval layer that lets a language model query the whole corpus.

From there it can assemble show recaps, cut clips and answer questions about what was said, where and how often.

The hard part was dealing with South African dialect and names. Place names, brand names and people’s names came out wrong, in the confident way transcription models get things wrong. “It would hear one thing and then actually write it in its own way that it’s actually understood,” he explained. “Whisper is actually brilliant and it’s highly accurate, but there’s always going to be those issues.”

His fix was unglamorous: for a couple of months he held the system’s hand, watching the headlines, listening to the shows, reading the transcripts and correcting them one term at a time into a lookup library. “That’s a manual process,” he said. He expects it to matter less as the models improve.

The feature with the most obvious life in it is the newest. Locl now records predictions made on air – by hosts and by callers – and goes back afterwards to check whether they came true.

“Who’s going to win this weekend’s game, and who made that call?” Trennery says. “And then after the fact we use AI to go and look up the actual result. Did this actually happen?”

The obvious application is politics. “If you look at what’s happening with the ANC and the local elections – we’re going to win this, or we’re not going to win this.