Smallest.ai Lands $13M to Build Conversational Voice AI Without Perceptible Delays
The startup is developing specialized voice models that process speech in real-time, mimicking natural human dialogue instead of relying on large language models for voice interactions.
Last verified:
Smallest.ai, an outfit established in late 2024, is attacking the voice-agent market from a different angle: rather than compressing large models, the team is engineering minimal architectures tuned exclusively for speech. The company just secured $13 million in its Series A funding round, with Seligman Ventures leading and Sierra Ventures and 3one4 Capital joining in. This injection pushes cumulative financing past the $21 million mark.
The core insight driving the startup’s strategy is that humans engage in overlapping conversation—listening, formulating responses, and speaking happen concurrently. Typical LLMs process sequentially: ingest input, compute, then output. According to Sudarshan Kamath, Smallest.ai’s founder and chief executive, this delay-prone architecture feels unnatural when mapped to voice. “If you think about how we are talking, I’m not giving you like a large clipping of my audio, and then you start thinking,” Kamath explained to TechCrunch. Even milliseconds of silence in a dialogue register as artificial.
Dual-Model Architecture for Voice Interactions
Smallest.ai’s technical approach centers on a two-tier system. A compact voice model handles domain-specific queries in real-time with near-zero latency, while a broader foundational model sits in reserve. When the compact model encounters a question outside its training scope, the system transparently escalates—mimicking how human support agents would research an unfamiliar issue. This hybrid design lets customer-service teams deploy specialized voice agents without engineering full LLM inference stacks.
The startup’s existing customer roster includes RingCentral and Truecaller, both established players in communications infrastructure. According to TechCrunch, Kamath indicated that customer-support platforms—both established companies and newer entrants like Sierra and Decagon—represent the addressable market. His pitch is practical: building world-class voice capabilities distracts teams from their core product.
Market Positioning Against Incumbents
Smallest.ai enters a crowded landscape. ElevenLabs dominates the voice-AI category, while Cartesia and Sarvam (which emphasizes non-English languages) stake regional claims. Unlike these competitors, which span use cases from podcast production to dubbing, Smallest.ai narrows its focus to synchronous customer interaction—a bet that specialization drives better real-time performance than generalist tools.
Why This Matters
The shift toward purpose-built voice models over adapted LLMs signals a maturation in the voice-AI segment. If Kamath’s thesis holds—that all voice applications eventually require dual-model stacks—the economics of voice support infrastructure shift. Support platforms that once might have built proprietary voice layers in-house will instead contract with specialists, reshaping vendor relationships in customer-service software. The question is whether Smallest.ai’s focus on latency and naturalness translates to measurable customer satisfaction and retention gains—metrics that will determine whether this funding round represents a sustainable category or a niche play.
Frequently Asked Questions
How does Smallest.ai's approach differ from shrinking large language models for voice?
Smallest.ai builds voice-specific models that process audio in real-time using simultaneous listening and response generation, rather than applying the sequential prompt-response cycle of traditional LLMs to speech.
What happens when Smallest.ai's model encounters an unfamiliar topic?
The system escalates to a full-scale foundational model, placing the caller on brief hold—a behavior designed to match human customer service workflows.
Who are Smallest.ai's competitors?
The company competes with ElevenLabs in voice AI, as well as Cartesia and language-specific players like Sarvam AI.