Smallest.ai raised $13 million in a Series A round led by Seligman Ventures, with participation from Sierra Ventures and 3one4 Capital. The financing brings the voice-artificial-intelligence startup’s total capital above $21 million and will support a product designed for real-time enterprise conversations.
Founded in late 2024 by Chief Executive Sudarshan Kamath, the company argues that natural voice interaction requires a different architecture from text chat. A conventional large model waits for a completed prompt before processing it, while human speakers listen, formulate an answer and sometimes interrupt at the same time.
Smallest.ai is developing a compact, specialized model intended to listen, reason within a defined knowledge domain and speak simultaneously. The company says this design produces virtually no response lag, a critical feature in customer-service calls where even short pauses make an automated agent sound mechanical.
The architecture does not try to make the small model answer everything. When a query falls outside its limited knowledge base, the system calls a larger foundation model and briefly puts the customer on hold while it obtains the information. Kamath expects future voice agents to combine a real-time conversational layer with a slower model for complex work.
The startup concentrates on voice-specific problems: accents, dozens of languages, interruptions and noisy environments. That focus differentiates it from broad foundation-model developers and from voice companies that use the same technology for dubbing, podcasts or media production rather than live enterprise calls.
RingCentral and Truecaller are among its existing customers. Smallest.ai competes with ElevenLabs, Cartesia and regional language specialist Sarvam, while customer-support platforms such as Sierra and Decagon could be buyers or could choose to build their own voice systems.
The commercial argument is that support-software companies should not divert engineering resources into becoming voice-model specialists. That thesis depends on measurable reliability, language coverage, latency, integration costs and the ability to hand a conversation between models without confusing the caller or exposing sensitive customer data.
Kamath says the goal is for users not to know whether they are speaking with a machine. That is an ambition, not an independently established result. The Series A validates investor interest and funds product development; the next proof will come from enterprise deployments, retention, unit economics and performance under real accents, noise and unpredictable requests.
