One of the problems with AI voice tools is the potential for your voice to be stolen by the AI that processes it, then used for training purposes or cloning.
A new innovation aims to add a further level of protection to prevent that.
Public open AI voice cloning and vocal interaction tools are succeptable to voice theft, giving AI Models the ability to clone your voice for their own purposes and also to steal information from your conversations.
With more secure models, you can protect data when it is stored and moving, but the hard part until now has been protecting it while the AI was actually using it. Deepgram now offers a solution for real-time voice AI.
Real-time AI infrastructure company Deepgram and Fortanix have embarked on a partnership that will enable enterprises to run voice AI in their own environment on their own terms while ensuring their most sensitive data is securely protected.
Deepgram will leverage Fortanix Confidential AI and NVIDIA Confidential Computing to add an additional layer of advanced security to self-hosted environments to ensure that its proprietary model weights, built on business-critical intellectual property, can be deployed while protecting against model theft or inappropriate use. This will enable increased voice AI adoption in highly regulated industries.
In radio, our voices are out there in public, so are always vulnerable to cloning. While this new development will not protect radio broadcasts and streams from being externally recorded and cloned, it will provide an extra level of security for real time radio streaming, multilingual translation and captioning as the stream is being sent.
Other organisations, such as those handling medical patient conversations, financial transactions, or classified information increasingly require that sensitive audio and AI model weights remain protected not only at rest and in transit, but also during active processing in their own environments. The increased level of protection enables organisations to build highly-secure real-time voice applications without sacrificing performance.
The on-premises solution runs Deepgram’s voice AI models with Fortanix Confidential AI on NVIDIA Confidential Computing-enabled GPUs, creating a hardware-isolated environment where both audio data and model weights remain encrypted and protected throughout active use, in contrast to online cloud services where the audio content may be exposed, especially in free online services such as AI note takers.
The solution opens the door to a variety of on-prem security-demanding voice AI applications: private, on-prem voice agents handling sensitive customer and patient interactions; enterprise-wide transcription layers that capture every call, meeting, and internal conversation for analytics, compliance, and search; and voice-enabled IT, operations, and service desk applications running entirely inside an organization’s secure perimeter.
It is good to see professional services getting serious about voice security.
Analysis by Steve Ahern
