As AI agents become more adept at addressing customer support issues, most individuals can still easily discern when they are interacting with a machine rather than a human.

Smallest.ai, a startup established in late 2024, is wagering that the next advancement in voice agents will arise not from accelerating large language models, but from deploying smaller, specialized models crafted for human dialogue. In essence, the firm’s goal is to make conversing with an AI agent just as seamless as chatting with a person.

To achieve this, it is working on a compact voice model that aims to replicate the way humans process information by listening, contemplating, and speaking concurrently.

“While I’m conversing with you, you’re already processing information, and you might interject if I speak for an extended period,” Sudarshan Kamath (shown left), founder and CEO of Smallest.ai, shared with TechCrunch, adding that this is precisely how the startup’s model is intended to operate.

To support this endeavor, Smallest.ai has secured $13 million in a Series A funding round, led by Seligman Ventures, with contributions from Sierra Ventures and 3one4 Capital. This infusion of funds elevates the startup’s overall funding to more than $21 million.

“The functioning of an LLM involves providing it with a complete prompt, after which it begins to process,” Kamath stated. While such latency is tolerable in a text chat, even a brief silence feels unnatural in a voice conversation. “If you consider how we communicate, I’m not sending you a lengthy segment of my audio, and then you start to ponder.”

The startup’s model operates as a real-time intelligence layer, facilitating natural discussions with customers on particular subjects, with nearly no response delay. However, if the model comes across a topic outside its narrow knowledge scope, Smallest.ai routes the inquiry to a large foundational model, momentarily putting the customer on hold to “research” the matter—just as a real person would do.

Kamath envisions that all AI agents will soon depend on two models: a small voice model for instantaneous interaction, and an “offline” LLM that is summoned as necessary to tackle intricate problems.

In contrast to large foundational models, Smallest.ai zeroes in specifically on voice-related subtleties, such as accommodating various accents, supporting a multitude of languages, and functioning efficiently in noisy settings.

Among the startup’s current clientele are companies within the voice sector, including RingCentral and Truecaller. Kamath noted that any customer support organization, including newer entrants like Sierra and Decagon, could be a potential client for the startup.  

When questioned why a well-capitalized AI customer support company wouldn’t develop its own voice model, Kamath replied that for customer support startups, excelling in voice capabilities can detract from their primary focus.

Smallest.ai competes with voice AI frontrunner ElevenLabs, as well as Cartesia and regional entities like Sarvam that cater to local languages.

While some rivals leverage voice AI for applications such as audio dubbing and podcast production, Smallest.ai dedicates itself solely to real-time conversational voice agents for its enterprise clients.

“Our aim is for our models to surpass the Turing test,” Kamath stated. “You should interact with our model without recognizing whether it’s AI or human. That’s the company’s singular objective.”