‍"It's lightning quick to get a report back from Resemble. It gives us a significant level of additional confidence that we're not barking up the wrong tree."

- Dominic Watt, Senior Consultant, JP French International

Meet Kristina and Dominic

Dr. Kristina Tomić is a Consultant and General Manager at JP French International (JPFI) and a member of MENSA. She has worked on more than 300 forensic cases, from speaker comparison to authenticity analysis and AI detection.

Prof. Dominic Watt is a Senior Consultant and Honorary Professor of Phonetics and Sociolinguistics at the University of York. He has spent 25 years working on forensic cases and helped establish the UK's standard for the field. In 2022, he received a Lifetime Achievement Award for his research.

Between them, they have decades of experience listening to recordings, studying spectrograms, and making judgments about what they hear.

Then AI-generated voices became realistic enough to raise a new question: could the recording itself be synthetic?

“Have you considered that this might be a deepfake?”

For years, JPFI’s forensic analysis started with the assumption that a recording was genuine, and speaker comparison focused on determining whether two recordings belonged to the same person. But as voice cloning became more realistic, the team had to consider a different possibility: what if the recording itself had been generated by AI?

By 2024, Dominic was increasingly hearing a new question from defense counsel: “Have you assessed the possibility that this is generated by a deepfake audio system?”

JPFI wanted to be able to answer that question before it became an issue in court. “If you’ve already done the pre-screening check, we can basically set that aside. It doesn’t take us by surprise,” Dominic said.

That raised a new challenge for the team as there was very little research on how specific acoustic parameters differ in AI-generated and genuine speech. Because deepfake technology evolves at a breakneck pace, cutting-edge findings are primarily shared at niche scientific conferences and it takes time until these become available in widely accessible journals. JPFI and the team also lacked access to extensive databases preventing them from directly comparing the spectral characteristics of the disputed audio within their own lab. They had to draw on their experience in speech and language science to assess whether a recording might be AI-generated and interpret the findings based on auditory-perceptual cues, linguistic inconsistencies, and known structural artifacts left behind by generative algorithms.

When JPFI needed an additional assessment of a disputed recording, they sometimes involved external experts which added considerable cost to the process, and those consultants often would not disclose their methods for proprietary reasons.

Before Resemble, Kristina said, “I would have to export to the experts outside of the company.” That made it harder for JPFI to fully stand behind a result they had not produced themselves.

Any new approach also had to meet the standards of forensic work. The UK courts had banned automatic speaker recognition evidence roughly fifteen years earlier over reliability concerns, a decision that still shapes how courts think about new methods.

For JPFI, that meant a deepfake detection tool couldn't simply produce an answer. it also needed to be validated, benchmarked, and defensible under cross-examination.

Screening comes first

JPFI now runs a synthetic-audio check before starting speaker comparison.

Traditional speaker comparison is about determining whether two recordings are from the same speaker. But if the recording itself was generated by AI, that analysis can miss the bigger question of whether the speech is authentic in the first place. So when a file could plausibly have been generated by AI, the team screens it before starting the speaker comparison. This includes voice notes, phone recordings, and other files where there is no accompanying video.

Several of the disputed recordings they have worked with were also not in English. Alongside its detection result, Resemble provides a transcription and translation, which the team has found genuinely useful.

The automated result is still only one part of the process. JPFI checks it against the sample from the known speaker and uses its own forensic methods to assess whether the speech has the characteristics of authentic human speech.

“We would also double check what it's telling us,” Dominic said, “using more human-based methods: are there features in the recording that have the hallmarks of authentic human speech?”

When the automated result and the manual analysis agree, that gives the team additional confidence. When they disagree, it tells them to look more closely.

Dominic describes it as a “belt-and-braces” approach: two different methods working alongside each other, giving the team another way to catch something one approach might miss.

Accuracy, tested twice

For JPFI, accuracy was the most important consideration when evaluating detection tools.

“I think we were aiming for accuracy. That’s the most important thing,” Kristina said.

Dominic had also seen Resemble’s system perform well in benchmarking trials, too. “I did see Resemble in benchmarking trials has been shown to be the best of the bunch. And that’s why we use it.”

But JP French wanted to test the system themselves. The team independently and repeatedly tested it using synthetic recordings created specifically to try to fool the system.

Some of those recordings took longer than usual to process. “It usually produces the results instantly,” Dominic said. “But with the recordings that we prepared, it had to think.” The system took longer, then returned the correct result.

The benchmarking results have also been useful when JPFI have had to discuss the system in court. In one recent case, Dominic needed the latest accuracy figures before cross-examination in an English court. “It’s just really very reassuring to see those benchmarking results,” he said. Without them, Dominic would only have been able to claim that the system was regarded as the best available. The opposing counsel did not pursue the line of questioning further.

A case where it changed the outcome

The biggest change has been what happens before speaker comparison begins.

In one case, a synthetic recording initially “sounded very like the suspect.” Resemble flagged it, giving the team a reason to investigate the recording itself before applying a protocol designed for authentic speech. That’s the point of the screening step: catching a potential problem before it shapes the rest of the analysis.

It has also changed how JPFI talks about these cases in court. Previously, disputed recordings could be sent to an external consultant whose methodology was not fully disclosed. Now, Kristina and Dominic can point to a named tool that has been independently benchmarked and has a public track record.

For JPFI, it gives the team another piece of evidence to consider alongside their own analysis.

Speech generation has advanced quickly enough that Kristina now sees a limit to what traditional speaker comparison can establish on its own.

“With the latest speech generation developments, since May 2026, Resemble AI Detect is possibly the only tool at our disposal for flagging AI generated content,” she said. “The voice clones have become so realistic that standard linguistic analysis may no longer make sense.”

She added that Resemble Detect has, on several occasions, redirected the team's analysis entirely, pointing them toward a different line of investigation from the one they started with.

Using AI detection in forensic work

Voice generation is moving quickly, and the standards for using AI detection in forensic work are still taking shape.

Dominic expects courts to become increasingly cautious about tools and methods that have not been properly validated. For JPFI, that makes the approach they have taken with Resemble increasingly important: use a validated and benchmarked tool alongside traditional forensic analysis, and be prepared to explain how the result fits into the wider evidence.

The team also doesn’t assume that better voice cloning makes every recording suspect. The goal is to identify the recordings that warrant a closer look, without making every piece of evidence harder to trust.

As the technology and the case law continue to evolve, JP French International expects that balance to keep changing too.