Voice AI Evolves as Key Infrastructure for Device Interaction
Voice AI is rapidly emerging as a fundamental layer of technology infrastructure, allowing users to engage with devices through natural conversations rather than traditional screen-based interfaces. This evolution is increasing the demand for robust compute power, extensive data storage, and strategic partnerships to support the underlying AI models that enable these conversations.
Deepgram Inc., which expanded its Flux model to support ten languages earlier this year, is among the companies addressing this demand. According to Nick Mann, staff technical program manager at Deepgram, the company requires substantial infrastructure to support both its internal research team and a growing customer base that numbers in the tens of thousands.
Mann highlighted the expertise of Deepgram’s research team, stating, “We have 45 of the smartest researchers you’ve ever seen who can go prospect and find amazing things. You know what they like to eat? Compute. Just tons of compute, tons of data.” This emphasis on computational resources underscores the increasing complexity and sophistication of AI development in the voice domain.
During a recent interview with John Furrier of theCUBE Research at Dell’s AI Leadership Symposium, Mann discussed the advancing voice AI infrastructure, detailing how Deepgram continually enhances its models and the ambition to run voice AI directly on personal devices. The session was part of a special broadcast by SiliconANGLE Media, aimed at exploring trends in AI technology and infrastructure.
From Research to Real-World Applications in Voice AI
With 13 years of industry experience, Deepgram has developed expertise in tuning AI models for challenging environments such as air traffic control and emergency medical services. Mann indicated that the company views each model release as a temporary milestone, each ripe for further improvement through retraining. “Let’s say it was trained 10 times to get as good as it was when we released it. What’s stopping us from retraining it on better data?” he posited.
Importantly, many of Deepgram’s AI models run outside the confines of its cloud infrastructure. A significant number of clients opt to host these models independently, and a recent partnership with Fortanix Inc. has opened opportunities for deploying private, on-premises voice AI solutions in regulated sectors. Mann noted that Dell Technologies Inc. provides the infrastructure necessary for Deepgram’s operations while also utilizing these models in its Customer Solution Center Innovation Lab.
This collaboration with Dell enhances Deepgram’s outreach to a diverse clientele, ranging from small enterprises to large corporations, thereby broadening the accessibility and application of voice AI technology. Mann pointed out that these implementations span from basic interactive voice response systems on compact Nvidia Corp. GB10 devices to extensive deployments across large-scale organizations.
However, Mann also emphasizes the need to expand voice AI capabilities beyond traditional data centers. He envisions a future where AI models can seamlessly run on the AI-enabled laptops that students carry to school. “I want our models on those, on the hardware, to run natively,” he elaborated. “Air-gapped. That’s where we want to get to. That’s where AI as a whole needs to start meeting people where they are and how they’re going to use it.”