OmniSpeech seeks additional seed funding to expand its real-time defense against voice deepfakes

This story was made possible through support from TEDCO, the Maryland Technology Development Corporation, which enhances economic empowerment growth through the fostering of an inclusive entrepreneurial innovation ecosystem. TEDCO identifies, invests in, and helps grow technology and life science-based companies in Maryland. Learn more at tedcomd.com.

As AI makes it easier to fake voices, tech is struggling to keep up with helping users determine what’s real.

Dozens of tools and startups have emerged to address the growing problem. Many of them are only trained to identify known deepfake algorithms, according to an NBC News report.

As voice-cloning technology rapidly evolves, those systems often struggle to detect new or unfamiliar data. More than a decade in the making, College Park, Maryland-based OmniSpeech is throwing its hat in the ring to compete, with plans to raise additional seed funding early this year to scale product development. 

What makes this startup different, according to CEO David Przygoda, is that its proprietary algorithm can identify synthetic voices from brand-new technologies. For instance, it can spot a deepfake from a tool’s data that it’s never trained on before. 

“It’s an uphill battle to continuously train new models and have them up to date on the latest technology,” Przygoda told Technical.ly. 

OmniSpeech’s AI Detect app is available on Zoom to scan calls in real time. The “zero-shot” AI model behind its app can spot deepfakes from previously unseen AI-voice generators by learning general patterns from large, diverse voice datasets. 

In 2024, the startup won an award from the Federal Trade Commission for its voice detection model. 

OmniSpeech’s first product uses machine learning and speech signal processing to eliminate background noise. It’s currently built into headsets for an unnamed major hardware company and can also be used in phones and other software applications.

As deepfakes get easier to build, detection gets harder

The spread of online deepfakes has surged in recent years. 

Cybersecurity firm DeepStrike reported an exponential increase in deepfakes between 2023 and 2025, rising from about 500,000 incidents to over 8 million. The report also found that voice cloning is among the most common forms of deepfake attack, with one in four adults saying they have experienced an AI-generated voice scam.

The ease of creating these fakes is driven by several factors. Scammers often need only a few seconds of audio to produce a voice clone. A magician who created a fake robocall mimicking former President Joe Biden circulated during a 2024 New Hampshire primary said it took just 20 minutes and $1 to produce.

“It’s a big problem for consumers, for enterprises and governments,” Przygoda said. 

OmniSpeech’s work traces back to with Carol Espy-Wilson, an electrical engineering professor at the University of Maryland and an expert in speech communication research. She founded the company in 2014 to translate her lab’s research on noise suppression into a market-ready product. 

New cash could help the team expand beyond Zoom for AI Detect, Przygoda said, including a Chrome extension that can identify deepfake voices on websites like YouTube, as well as a standalone application.

“Right now, we’re focused on strategic partnerships and integrating our technology within existing platforms,” Przygoda said.