There is a reason many organizations in healthcare, finance, and the government sector, along with other regulated businesses, have moved cautiously on voice AI. It is not because the technology lacks promise, but because voice data is often among the most sensitive information an enterprise handles. It’s not a use case or underlying technology question, but a security one. Protecting voice data has generally been harder once it is actively being processed.
When a clinician dictates patient notes, when a financial professional discusses a transaction, or when a government analyst speaks about sensitive operations, that audio is not just another data stream. It is regulated, high-value information. Sending that data through external cloud infrastructure or running it on shared systems introduces enough security and compliance uncertainty to preclude adoption.
That doesn’t have to be the case. Deepgram, Fortanix, and NVIDIA are trying to solve that challenge with a new on-premises deployment model built around confidential computing.
The key issue is not data at rest or data in transit. Enterprises already have mature tools for encrypting stored files and protecting information as it moves across networks. The challenge is data in use — when audio is actively being processed by an AI model. That has typically meant the data is decrypted in memory, where it could be exposed to the host OS, privileged administrators, or attackers who have gained access to the underlying infrastructure. That’s a barrier to adoption for regulated industries.
The new Deepgram-Fortanix-NVIDIA stack is designed to close that gap. Deepgram provides the voice AI models; Fortanix brings its Confidential AI platform, which handles attestation, trust, and protection for models and data in use; and NVIDIA adds its Confidential Computing-enabled GPUs, which create Trusted Execution Environments that isolate the workload from the underlying infrastructure and keep both audio data and model weights protected. Put simply, the three companies are offering a jointly engineered on-premises stack that lets enterprises run real-time voice AI while reducing exposure to surrounding environments.
Enterprises want assurance that sensitive audio remains protected while their model is actively working on it. At the same time, Deepgram and other model owners want to protect proprietary model weights from extraction or misuse. This partnership is designed to solve both problems.
The collaboration changes how regulated businesses look at voice AI. Healthcare organizations can run voice agents for patient interactions or clinical documentation inside their own environments. Financial institutions can capture and analyze calls for compliance and operational purposes without routing sensitive audio through third-party cloud infrastructure. Government and defense environments gain a more viable path to voice-enabled workflows that align with strict sovereignty, residency, and confidentiality requirements.
Of course, there are also broader enterprise applications, such as private transcription, internal analytics, and voice-enabled IT and service operations running entirely within an organization’s secure perimeter.
Security requirements around AI are becoming more demanding, voice interfaces are becoming more useful. The appeal of this partnership is, in fact, not just stronger protection, but the combination of stronger protection with the real-time performance voice applications still require. In other words, it makes voice AI more operationally acceptable in the environments that have historically had the strongest reasons to hold back. Organizations that have been interested in voice AI but haven’t been able to reconcile it with their security postures may have a new path forward with this confidential-computing-based design.
Edited by
Erik Linask
