Conversational radio in South Africa has long been a vast source of news that was impossible to index. The Locl platform was created to change this situation. It tracks live broadcasts from the country's largest English-language conversational stations, transcribes all content, indexes it, allowing users to find any name, brand, or topic and see all mentions on air.
Among the stations whose content is processed by Locl are 702, Cape Talk, Power FM, and SAfm, with plans to include platforms in other languages in the future.
The platform's creator, Sean Trennery, told TechCentral on Friday that his interest lay not so much in news segments as in listener calls, interviews, and discussions—the part of radio that was never indexed and disappears immediately after broadcast. He noted that podcast feeds provide entire shows but do not allow for finding specific moments.
Trennery has about 25 years of software development experience, but he developed Locl without writing the code himself. He stated that the platform is built entirely on Claude Code, and they are currently at a stage where they are no longer considering the code itself.
Creating the platform took three to four months, and work on it was done in his free time outside of his main activities. Trennery explained that ideas came to him during the day or while jogging, after which he passed them to AI, which then iterated the process and generated a pull request for his review and merging.
How the system works
For each station, a listening tool is launched that records the live broadcast in short half-minute segments. Each such segment is processed by an open-source speech recognition model, Whisper, and the resulting text enters a massive index that Trennery calls. Above this index is an extraction layer that allows the language model to query information from the entire data corpus.
This enables the creation of show summaries, clipping, and answering questions about what exactly was said, where, and how often. The main difficulty was correctly recognizing the South African dialect, as well as names and places and brands. Transcription models often make mistakes, confidently providing incorrect information. Trennery explained that the model could hear one thing but record something completely different that it thought was understandable.
He admitted that although Whisper is very accurate, problems remain. His solution was manual: for several months, he manually monitored headlines, listened to broadcasts, read transcripts, and corrected terms in a reference library. He expects that as models improve, this manual work will become less necessary.
A newest feature is the ability to record predictions made on air by both hosts and calling listeners, followed by checking their accuracy. Trennery gives the example of a question about the winner of an upcoming match, which is then checked using AI against the actual result.
The obvious application of this feature is politics. For instance, one can track predictions regarding the ANC and local elections to determine which statements were true and which were not.
In addition, the platform includes a number of additional products: a daily digest of everything discussed on all stations (available for reading or listening via RSS), summaries of individual shows, alerts for brands or names, sentiment analysis (showing whether an entity is mentioned positively or negatively), and tracking which advertisers buy the most airtime.
Trennery created a hub around the World Cup that was taking place and another for the upcoming local elections. He also presents this system as an element of a news workspace.
Most of the service is available for free, and Trennery intends to maintain this model. The commercial component is aimed at brands that require a full archive, workspaces, and reports. Copyright issues arise when processing the live content of commercial radio stations. Trennery stated that he has studied these legal aspects and believes he is fine because he uses open streams that broadcasters already publish for their listeners.
Three levels of access
According to the website, Locl operates on three levels. The free level requires no payment and includes a daily digest, preliminary search, and podcast feed. The Pro package costs 199 Rand per month including VAT and provides unlimited search and chat, five email keyword alerts, audio clips, and ad analytics—an overview of which brands buy the most airtime. The highest level, Team, costs 999 Rand per month for a group of up to eight users and includes unlimited alerts, shared clips, chats, and search. Each account starts with a seven-day free trial of Pro.
The future development of the platform, according to Trennery, lies in expanding to television news channels, as audio content is already present there. There is also consideration for supporting other languages and markets in English-speaking countries, although he has not yet explored these areas.