If you have ever conducted an interview, you know the moment when the dynamic shifts. The recorder clicks off, and you’re left with an hour of audio, messy transcripts, and pressure to move quickly without sacrificing accuracy. The value is obvious. The workflow, less so.
Interviews are among the richest sources of original reporting and among the most demanding to turn into publishable journalism. A single hour-long conversation can generate multiple stories, follow-ups, and sidebars. All of them depend on careful transcription, verification, and editing. For editors, the challenge is not just speed, but maintaining consistency and control across that entire workflow.
To explore how automation might support this process without compromising editorial standards, we partnered with Atex on the JEDI Project. The aim is straightforward: help turn raw interview recordings into structured, high-quality article drafts that fit naturally into existing newsroom workflows.
The project is supported by the FAIR EU Fund. The FAIR initiative is promoted by the Italian Ministry of University and Research (MUR) and funded by the European Union through NextGenerationEU.
What JEDI Is
JEDI is not a simple transcription tool, and it is not meant to replace journalists or editors. Its role is to take on some of the repetitive work that follows an interview – work that often slows publication and pulls attention away from higher-value editorial tasks.
Alongside a verbatim transcript, the system produces a structured article draft that is intended to be close to publishable, while remaining fully traceable to the original interview. Editorial judgment is still essential: the output is meant to be reviewed, questioned, edited, and shaped by people.
One of the trickier aspects of interviews is how non-linear they tend to be. Interviewees jump between topics, circle back to earlier points, and often expand on ideas much later in the conversation. JEDI is designed to handle that reality by identifying themes across the entire recording and assembling topic-focused sections from different parts of the interview.
This approach also makes it possible to produce multiple articles from a single long interview, with each draft focused on a distinct subject.
In practice, this means that a long interview about climate policy can produce one article on regulation, another on industry response, and a third focused on local impact – without an editor having to manually stitch those pieces together.
At the moment, the system supports recordings of up to 60 minutes and typically produces drafts in the 2,000–3,000 word range. Drafts can be aligned with publication-specific style guidelines, allowing them to match an outlet’s tone and structure instead of defaulting to a generic format.
While development has focused primarily on English, we also tested the system on German, Czech, and Slovak interviews and were happy with the results.
Traceability, Verification, and Editorial Control
From the very beginning, it became clear that accuracy alone was not enough for newsroom use. Editors need to understand where information comes from and why it appears in a draft. That requirement shaped several key design decisions.
Every paragraph in a generated article is linked back to the original transcript and audio. This makes it easy for editors to quickly verify claims, check phrasing, and confirm context without digging through raw recordings.
The system also includes automated fact-checking. Factual claims are first checked against the interview transcript itself and, where relevant, compared with publicly available information. This does not replace editorial fact-checking; it simply flags potential issues earlier, when they are cheaper and easier to fix.
From Audio File to Editable Draft
Behind the scenes, the JEDI Project relies on a modular agentic architecture that combines speech processing, large language models, and editorial logic. For newsroom users, however, the workflow can be understood in three straightforward stages:
1. Transcript Cleanup Stage
The process starts by improving transcript quality:
- Merging fragmented sentences caused by natural speech patterns
- Smoothing out interruptions
- Correcting grammar and clarity issues
All changes are checked against the original transcript to ensure that meaning is preserved and nothing new is introduced.
2. Article Drafting Stage
Once the transcript is cleaned, an article draft is generated:
- Proposing headlines
- Writing an introduction
- Dividing the content into topic-based sections
- Assigning headings
- Breaking long or complex answers into smaller, more readable parts, sometimes using generated follow-up questions to clarify structure
If a draft exceeds the target length, less essential passages are selectively condensed while preserving the main points.
3. Review and Verification Stage
In the final stage, the focus shifts to refinement and consistency:
- Normalizing terminology
- Clarifying temporal references
- Reviewing the article sentence by sentence
- Linking each factual claim back to the source interview, with direct access to both transcript and audio, so editors can verify details efficiently before publication
How New Approaches Are Evaluated
Evaluation plays an important role in developing the JEDI Project. Instead of relying only on automated metrics, our team regularly compares different versions of the same article during the testing phase, such as drafts produced by different language models, configurations, or system updates.
Using an automated “LLM-as-a-judge” approach, drafts are compared side by side and assessed against clear editorial criteria (clarity, consistency of voice, handling of interview responses, terminology). The evaluation tool then selects the stronger version and provides an explanation as to why it did so.
While automation helps with the comparison, journalists review a subset of the results to validate the outcomes and provide feedback. This combination of automated comparison and human review makes it easier to understand not just which version of the system performs better, but why.
Where the Project Stands Today
The first phase of the JEDI Project is complete, and the system is already usable in real editorial workflows. Current work focuses on improving accuracy, expanding multilingual support, and increasing transparency for editors.