Deepgram, Fortanix and NVIDIA Team Up to Bring Confidential Voice AI to Regulated Industries

As enterprises race to deploy AI-powered voice applications, one obstacle continues to slow adoption in highly regulated sectors: protecting sensitive data while it is actively being processed.

Deepgram is aiming to address that challenge through a new partnership with Fortanix that combines confidential computing technology from NVIDIA to create a secure deployment model for enterprise voice AI. The collaboration is designed to help organizations run speech recognition, transcription, and voice agent workloads inside their own environments while keeping both sensitive data and proprietary AI models protected during live inference.

The announcement targets industries where privacy and compliance requirements often prevent organizations from fully embracing AI technologies, including healthcare, financial services, government agencies, and critical infrastructure operators.

Traditional security controls focus on protecting information while it is stored or transmitted. However, data often becomes exposed during processing, creating a potential attack surface for cybercriminals, insider threats, or infrastructure compromises. The new offering seeks to close that gap by leveraging confidential computing environments that isolate workloads and maintain encryption while AI systems are actively operating.

"Voice often contains the enterprise's most sensitive data — patient conversations, financial transactions, classified briefings. As real-time voice becomes the primary interface of the enterprise, organizations will need to deploy voice AI in the environments that their security, confidentiality, and regulatory requirements demand," said Scott Stephenson, CEO and Co-Founder of Deepgram.

The solution combines Deepgram's speech and voice AI models with Fortanix Confidential AI running on NVIDIA GPUs that support confidential computing capabilities. The architecture creates a protected execution environment designed to prevent access to sensitive audio streams and model weights, even from privileged system administrators.

Security experts increasingly view model protection as a growing concern as organizations deploy proprietary AI systems across hybrid and on-premises environments. Beyond safeguarding customer data, enterprises must also defend the intellectual property embedded within AI models themselves.

"AI labs and model owners face enormous pressure to commercialize and distribute their AI models widely, but doing so can create security risks that threaten the very intellectual property they invested huge amounts of time and money to create," said Anand Kashyap, CEO and co-founder of Fortanix. "Fortanix Confidential AI changes that by providing model owners a secure, proven path to unlock new markets and enterprise customers without exposing their most valuable IP."

The technology could support a broad range of security-sensitive applications, including private customer service voice agents, enterprise-wide transcription platforms, healthcare documentation systems, and voice-enabled IT operations tools. Organizations operating under frameworks such as HIPAA, GDPR, and data sovereignty regulations may find confidential computing increasingly necessary as AI adoption expands.

The announcement also highlights a growing trend across the AI industry. As enterprises move beyond experimentation and begin deploying production-grade AI systems, demand is shifting toward technologies that can secure data not only before and after processing but throughout the entire AI lifecycle.

"Real-time voice AI helps industries scale efficiency and safety — empowering healthcare providers to update records more quickly, letting technicians repair equipment without putting tools down to check manuals — but sensitive data needs to be secure throughout processing," said Justin Boitano, vice president of Enterprise AI Platforms at NVIDIA.

As voice interfaces become more deeply integrated into business operations, confidential computing may emerge as a foundational security requirement rather than a specialized capability. For organizations balancing innovation with compliance, the ability to run AI models inside protected execution environments could become a critical factor in determining how quickly advanced voice AI technologies are adopted across regulated industries.