Enabling Secure Enterprise Voice AI While Protecting Proprietary Models and Sensitive Data
Customer Profile
Deepgram is a leading voice AI foundation-model company trusted by enterprises and developers worldwide for speech-to-text, text-to-speech, and voice-agent applications. Its advanced models are built on a foundation of significant proprietary research, training and optimization, making them valuable intellectual property and central to the company’s market leadership.
Products
Fortanix Confidential AI
Challenge
Protecting Deepgram’s Model IP
Deepgram’s models, including weights, parameters and architecture, represent years of innovation and substantial investment. Deploying frontier models into third-party environments, whether on premises, in AI Factories, or across NeoCloud providers, introduces significant business risk. Model owners must consider that:
- Model weights could be copied or extracted
- Proprietary logic could be reverse-engineered
- AI assets could be reused without authorization
Securing Highly Sensitive Voice Data Across Enterprises
Voice AI often operates on some of the most sensitive information an organization handles. Depending on the industry, this could include:
- Patient histories and clinical conversations
- Banking calls and financial details
- Customer support interactions
- Internal meetings and confidential discussions
- Government or security-related communications
Fortanix Solutions
Deepgram addresses these challenges with Fortanix Confidential AI, enabling the secure deployment of proprietary voice AI models across enterprise environments without exposing model IP or sensitive customer data. Fortanix Confidential AI combines confidential computing, attestation, and secure key management to protect both models and data while they are in use.
In-Memory Encryption
Models and voice data remain protected during execution
Cryptographic Attestation
Only verified, trusted environments can run workloads
Secure Key Release
Model weights decrypt only under approved conditions
Zero Trust for Admins
Even privileged users cannot access protected workload data
Benefits
Deepgram’s model weights remain protected during execution, helping prevent theft, replication, reverse engineering and unauthorized access
With Fortanix, Deepgram can support self-hosted, on-premises and controlled enterprise AI deployments that would have previously been too risky from an IP perspective.
With proprietary models and internal data secured, advantages abound: healthcare organizations can securely transcribe patient interactions, financial institutions can responsibly analyze customer calls, and government agencies can deploy voice AI while maintaining operational control.
By reducing security concerns around runtime exposure, organizations can launch their voice AI initiatives to production faster and with far less friction.
Fortanix replaces implicit trust with cryptographic verification. That means: Environments must prove their integrity before execution. Keys are released only under approved conditions. Runtime exposure is minimized. Security controls are measurable and enforceable
