Deepgram, a voice-recognition firm, has partnered with security-software company Fortanix and chipmaker Nvidia to let regulated industries deploy voice AI without sacrificing data privacy. The collaboration brings together Deepgram’s speech-to-text engine, Fortanix’s confidential-computing platform, and Nvidia’s graphics processing units. The goal is to keep sensitive audio data encrypted even while it’s being processed — a requirement for sectors like healthcare, finance, and government.
Why regulated industries needed this
Voice AI has been slow to catch on in highly regulated environments. Laws like HIPAA in the U.S. and GDPR in Europe demand that personal data stay protected from end to end. But traditional cloud-based voice processing often exposes raw audio to the service provider, creating compliance headaches. Deepgram, Fortanix, and Nvidia say their combined approach closes that gap. Audio stays inside a hardware-backed trusted execution environment, or TEE, that encrypts data in use, not just when it’s stored or transmitted.
How the technology stack works
Fortanix is the layer that enforces that encryption. Its confidential-computing platform runs on Nvidia GPUs, which provide the compute muscle for Deepgram’s neural networks. The voice data never leaves that protected enclave. Deepgram’s model processes the speech, outputs a transcript, and discards the raw audio — all within the encrypted zone. The company says the setup meets the privacy standards that auditors look for in, say, a hospital call center or a bank’s customer-service line.
What this means for adoption
The partnership removes one of the biggest barriers to voice AI in compliance-heavy fields: the fear that a third party could access private conversations. Companies that had to do all speech processing on-premises, or skip it altogether, now have a cloud option that keeps data secure. Fortanix’s involvement adds attestation — a way to prove to an auditor that the encrypted processing is real. Nvidia’s GPUs handle the heavy lifting without exporting the data to a less secure location. Deepgram says early tests show no significant drop in accuracy or speed compared with its standard cloud service.
What comes next
The three companies are offering the integrated solution now, but they haven’t named the first customers. A regulatory nod is often the bottleneck in these industries, and each deployment will likely need its own compliance review. The next real test will be whether a hospital or a bank actually puts voice AI into production this way — and whether regulators accept the attestation as proof of privacy.
