Smallest.ai's $13M bet: voice AI's breakthrough isn't faster LLMs but small models that think while they listen

Smallest.ai's $13M bet: voice AI's breakthrough isn't faster LLMs but small models that think while they listen

The dominant strategy for voice AI has been to take a large language model, bolt speech input and output onto it, and optimize until the pauses shrink. ElevenLabs promises "sub-second responsiveness" across more than 70 languages on its conversational AI platform 1. Smallest.ai is raising $13 million to bet that approach is architecturally wrong for live conversation, and that real-time speech requires a purpose-built model rather than a faster LLM.

That thesis matters because it splits voice AI from the LLM arms race. If Smallest.ai is right, the metrics that determine winners in this market are not parameter counts or benchmark scores but conversational latency and the ability to know when to stop talking and ask for help. Voice AI would develop its own competitive stack, optimized for inference speed and conversational flow rather than the reasoning depth that drives general-purpose model development. The company, a startup founded in late 2024, has raised the $13 million in a Series A led by Seligman Ventures, with participation from Sierra Ventures and 3one4 Capital, bringing total funding past $21 million 2. Its existing customers include RingCentral and Truecaller 2.

The architectural difference is specific. Large language models process prompts sequentially: the user delivers complete input, then the model generates output. "The way an LLM works is you give it an entire prompt, and then it starts thinking," founder and CEO Sudarshan Kamath told TechCrunch 2. TechCrunch reports that this latency is acceptable in text chat but feels unnatural in a voice conversation, where even a short pause signals to the listener that they are talking to software 2. Smallest.ai's model is designed to listen, think, and speak concurrently, the way humans do in conversation, handling accents, dozens of languages, and noisy environments with near-zero response lag 2.

The handoff is where Smallest.ai diverges from every competitor named in its space. When the small model encounters a question beyond its knowledge, it escalates to a large foundational model and places the caller on hold to "research" the issue, the same way a human agent would consult a colleague 2. Kamath believes all AI agents will eventually adopt this two-tier design: a small voice model for real-time interaction and what he calls an "offline" LLM called on demand for complex problems 2.

That is a fundamentally different bet from what ElevenLabs is making. ElevenLabs positions its conversational AI as a single platform grounded in a customer's knowledge base, adapting to questions and edge cases in real time 1. Its product description centers on one integrated agent, not a split between a fast voice model and a slower reasoning model. TechCrunch describes ElevenLabs as a voice AI leader and names Cartesia as another competitor, noting that while some companies in the space apply voice AI to audio dubbing and podcasting, Smallest.ai focuses strictly on real-time conversational voice agents 2.

The competitive question for builders and investors is whether separation beats integration. A single-model approach keeps the conversation continuous but carries the latency overhead of a large model on every interaction, including simple ones. The two-model approach delivers near-zero latency for routine queries but introduces a pause whenever the small model hits its knowledge boundary. In a short support call, that escalation might happen once and feel natural. In a complex technical discussion, it could happen repeatedly, and each hold erodes the experience. Whether that trade-off holds up under real call volume is something only RingCentral and Truecaller's deployment data can answer.

Kamath positions any customer support company as a potential buyer, arguing that for those companies, becoming "extremely good at doing voice is a distraction from their core business" 2. TechCrunch names newer support companies including Sierra and Decagon as falling in that category 2.

"We want our models to break the Turing test," Kamath said 2. The field is converging on that target from different architectures. Whether the two-model handoff arrives first depends on something a Series A announcement cannot settle: does the pause feel like collaboration, or does it feel like computation?