April 2, 2026

Purdue ECE researchers to present Spanish deepfake speech detection paper at ICASSP 2026

The research is among the first to focus on Spanish deepfake speech at this scale. While much of the existing work in synthetic speech detection has centered on English and Chinese, the Purdue team created a large Spanish dataset designed to help researchers detect fake audio and identify which system produced it.
Young woman with long brown hair smiling warmly against a plain beige background. She wears a black top, conveying a friendly and approachable tone.
Maria Risques

Purdue University researchers will present new work on detecting AI-generated Spanish speech at the 2026 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) in Barcelona this May. The paper, “Detecting and Attributing Synthetic Spanish Speech: The HISPASpoof Dataset,” continues Purdue ECE’s work in deepfake speech and audio forensics.

The research is among the first to focus on Spanish deepfake speech at this scale. While much of the existing work in synthetic speech detection has centered on English and Chinese, the Purdue team created a large Spanish dataset designed to help researchers detect fake audio and identify which system produced it. The dataset includes real speech from six Spanish accents and synthetic speech generated by six text-to-speech systems.

The work was led primarily by Maria Risques, a student from Spain working in the lab of Edward J. Delp, the Charles William Harrison Distinguished Professor in Purdue’s Elmore Family School of Electrical and Computer Engineering. Delp is a co-author on the paper along with Risques, Kratika Bhagtani and Amit Kumar Singh Yadav.

“One of the most exciting parts of this project was helping create a dataset that focuses on Spanish, a language that has not received the same attention in deepfake speech research,” Risques said. “We hope this work gives researchers a stronger foundation for building tools that can better detect synthetic audio across languages and accents.”

As voice cloning tools become more realistic and easier to use, researchers are working to build better ways to spot fake speech that could be used for scams, impersonation or misinformation. The Purdue team found that detection systems trained on English do not reliably carry over to Spanish, showing the need for language-specific tools. Training on the HISPASpoof dataset significantly improved performance on Spanish synthetic speech.

“As synthetic speech tools become more powerful and more accessible, it is increasingly important to develop forensic methods that work beyond English,” Delp said. “This research helps address that gap by creating tools and data for Spanish, one of the most widely spoken languages in the world.”

Delp says the work is continuing by developing new methods and datasets for Italian.

ICASSP 2026 - the flagship conference of the IEEE Signal Processing Society - takes place from May 6 – 8 in Barcelona, Spain. It is the world’s largest and most comprehensive technical conference dedicated to signal processing.