Anyone who has done an interview for work knows the annoying truth: the hard part doesn't end when you stop recording. It starts. Hours of audio need to be played back, transcribed, cleaned up and shaped into something that actually reads well. For journalists and content teams working with tight deadlines and even tighter budgets, that process eats up time that could be spent on the next story.
That's the problem Atex set out to solve with its JEDI Project, an AI-driven system built to convert raw interview audio into near-finished articles, with only a light touch of human review needed at the end.
A lengthy interview can easily require several hours of transcription and drafting before a story is ready for editing. Those are valuable hours that journalists could otherwise spend investigating stories, speaking with additional sources or pursuing new angles.
JEDI was created with a simple goal: reduce the manual effort involved in turning interview audio into an article draft, allowing journalists to focus more on reporting and less on administrative tasks.
Many AI tools stop once they have converted speech into text. JEDI takes the process several steps further.
The platform can process up to 60 minutes of interview audio and generate a structured article draft of approximately 2,000 to 3,000 words. Rather than producing a raw transcript, it creates a narrative draft organized into logical sections that editors can review, refine and publish.
The workflow includes three key stages:
The system improves readability by correcting grammatical errors, resolving fragmented sentences and organizing dialogue while ensuring that changes remain faithful to the original recording.
JEDI identifies key themes within the interview and builds a structured draft with clear sections, helping transform a conversation into a coherent story. It can also suggest follow-up questions where additional context may strengthen coverage.
Accuracy remains central to the process. JEDI checks terminology and formatting while enabling factual claims within the draft to be traced back to the original audio source. This helps integrate fact-checking into the content creation workflow from the start.
The last stage is where JEDI focuses on getting the piece publication-ready. It smooths out formatting, checks that terminology stays consistent throughout, and reviews sentence structure for readability. More importantly, it pulls out every factual claim in the draft and checks it against the original transcript.
That traceability is a big deal for editors. Instead of digging back through an hour-long recording to confirm a quote or a figure, they can see exactly where in the source material a claim came from. It's a system built for accountability rather than one that just hopes the output is accurate.
Rather than relying purely on automated scoring, the Atex team compares different versions of the same article side by side, testing different language models, settings and system updates against one another. They also use an “LLM-as-a-judge” approach, where drafts get evaluated against clear editorial standards like clarity, consistency of tone and how well interview responses are handled. That feedback loop is what's driving the ongoing refinement of the tool.
Behind the scenes, JEDI runs on a mix of speech processing, large language models and editorial logic layered together. For the newsroom staff actually using it day to day, though, the experience boils down to three simple steps: get a clean transcript, generate a structured draft and review a fact-checked, publication-ready article.
The project is part of Atex's win in the FAIR EU Fund's Spoke 5 initiative, which is focused on high-quality AI, and it's being developed alongside partners including Geneea. Atex is currently running a beta program that gives newsrooms full access to the tool, direct support from the technical team and a real say in how the product develops going forward.
The bigger idea here isn't to replace journalists. It's to remove the drudgery of transcription and first-draft writing so the humans on the team can focus on what only they can do: exercising editorial judgment and getting the story right.
