Redefining Voice AI Efficiency: A Q&A with Sudarshan Kamath, CEO and Founder of Smallest AI

I’m Sudarshan, founder and CEO of Smallest AI. Before Smallest AI, I worked in computer vision and self-driving, including time on Bosch’s autonomous-vehicle team. I started in workplaces where AI models had to run in real time on specific hardware and simply couldn’t afford to be wrong. That experience is where the whole idea for Smallest AI came from. Most of what goes into a giant AI model is noise. Strip it out, train something small and specialized instead, and you can match a much larger model’s performance at a fraction of the size and cost.

I believed in that approach enough to buy the domain smallest.ai in 2022, two years before there was actually a company behind it. It became real once I moved into voice technology and found that every voice agent I tried was slow and obviously a machine.

Today Smallest AI is a voice-first research lab built entirely from small, specialized models rather than one giant one. Pulse for speech-to-text, Electron as our small language model, Lightning for text-to-speech, and Hydra, our native speech-to-speech model. That efficiency is what lets enterprises run real customer conversations on us at volume, affordably.

When we founded Smallest AI, we made a deliberate choice to solve one genuinely hard problem, rather than try to build everything. We developed real-time, human-quality voice that holds up on messy real inputs, not just a controlled demo with a clean accent and no background noise. Anyone can win a demo, but almost no one survives production where a strong accent, background noise, an interruption, or half a second of lag instantly gives the machine away.

That’s the lesson I can share with other leaders. If you’re up against bigger, better-funded competitors, don’t try to match their breadth. Find the narrow place where being small and specialized is actually an advantage, and be undeniably the best there. For us, that meant getting our models onto public benchmarks so customers could verify our claims themselves, rather than just taking our word for it, and building real relationships with the early customers who had that exact painful problem.

It’s important to know where your strength actually is, go deep there instead of wide, and lean on your first customers to prove it. Depth in the thing that beats breadth every time.

One of the biggest trends in voice AI is the drive by customer-facing businesses to maintain and develop consumer trust. Developing trust requires being there when your customer needs you and whether what you’re saying holds up. As voice AI starts sounding like a real conversation instead of a menu, the bar on trust gets higher. Now, the more human AI sounds, the more a business is on the hook to back that up with real performance, not just a polished demo.

Availability is where this matters most. A huge share of customer issues today sit unresolved simply because no business can staff enough people around the clock to address them. Someone calls their hospital at 2:00am about their medication, and there’s no one to pick up until morning. Voice AI’s real value isn’t just efficiency, it’s being genuinely available for the moments that can’t wait, without feeling cold or scripted.

That combination, always there and still approachable, is what builds trust. Businesses get there by staying transparent about what the technology can and can’t do, and by treating gaps honestly with customers instead of pretending everything works, so trust holds even in the moments that matter most.

Two big ones. The first is treating voice AI like a chatbot with a voice bolted on, when it’s really a real-time systems problem. Teams wire together speech-to-text, a language model, and text-to-speech, plug in a telephony provider, and assume they’re done. Every handoff in that chain adds delay, and once you’re routing through four or five services, callers feel it. They start talking over the agent, or they just hang up. Voice AI lives or dies on turn-taking, and most teams don’t account for that timing until it’s already broken in production.

The second one hits the technical crowd hardest, especially CTOs. They obsess over benchmark scores. Everyone wants one number to rank voice models, so they chase MOS or an LLM-as-a-judge verdict, and those numbers barely track what a real person feels in a live conversation. I’ve heard two voices post almost identical naturalness scores and sound nothing alike where it counts. I’ve watched a model win an “emotion” benchmark and then sound robotic the second a human listens, because it acts excited in every sentence instead of the one line that needed it.

What the scores miss is fit. A bright, upbeat voice on a bereavement line reads as tone-deaf, however high it scored. A slow, careful voice on a quick task feels wrong for the job. The real question is whether a voice sounds like the right human for that exact moment, and no leaderboard measures that yet. So teams optimize the number and ship it, then wonder why customers still don’t trust the experience. Closing that gap, between the score and the moment, is where we spend our time.

In a crowded market, don’t ask customers to take your word for it, give them a way to check for themselves. When everyone’s messaging sounds the same, independent proof cuts through in a way persuasion never will. We learned this by shifting away from telling people about how good we are, and pointing them at public, third-party leaderboards where they could see the results themselves. That single shift did more for our credibility than any amount of marketing copy, because it moved the burden of proof off us and onto something the customer controlled.

The second piece is understanding that modernizing the customer experience isn’t only a technology problem, it’s a relationship one. Early customers rarely choose you because the product is polished, they choose you because they trust the people behind it. So don’t hide behind a slick interface, get in front of your customers directly, and pay attention to the gap between what they say they need in a meeting and what’s actually causing them pain. That gap is usually where the real opportunity sits, and closing it is what turns a one-time buyer into someone who sticks with your brand for the long run.