Human-Sounding Voice AI Pulls Smallest.ai Into $13M Race

$13 million is now riding on Smallest.ai voice AI making phone agents sound human enough that callers may not immediately know they’re talking to software. The startup has raised a Series A led by Seligman Ventures, with participation from Sierra Ventures and 3one4 Capital, bringing total funding to over $21 million, according to TechCrunch.

Smallest.ai’s bet is blunt: the next jump in AI calling won’t come from simply making bigger LLMs respond faster. It will come from smaller, voice-specific models built for the rhythm of human conversation.

$13 million Series A puts Smallest.ai voice AI on the latency problem

Smallest.ai, founded in late 2024 according to TechCrunch, is developing a small voice model that can listen, reason and speak at the same time. That matters because phone calls expose latency instantly. A pause that feels fine in text chat can make a voice agent sound fake before it finishes a sentence.

Founder and CEO Sudarshan Kamath framed the problem as a mismatch between how LLMs process prompts and how people actually talk.

“The way an LLM works is you give it an entire prompt, and then it starts thinking,” Kamath said.

He contrasted that with real conversation, where people begin processing before the other person stops speaking.

“While I'm speaking to you, you're already thinking, and you might interrupt me if I talk for too long,” Kamath told TechCrunch.

The company’s model is designed to act as a real-time intelligence layer for customer conversations on specific topics, with what TechCrunch describes as virtually zero response lag. If the system hits a topic outside its narrow knowledge base, Smallest.ai hands the query to a large foundational model and briefly puts the customer on hold to “research” the issue.

That architecture gives Smallest.ai voice AI a clear product thesis: keep the live call flowing through a specialized model, then call on a larger model only when the question gets complex.

TechCrunch identifies Kamath as founder and CEO. Separate supplied reporting from Inc42 names Akshat Mandloi as a co-founder and gives a different founding year, 2023. The supplied materials do not reconcile that discrepancy. Valuation and headquarters were not disclosed in the TechCrunch report.

The two-model phone stack is the real pitch

Kamath expects AI agents to move toward a two-model setup: a small voice model for the live interaction, plus an “offline” LLM for harder problems. That’s the most important technical claim in the raise.

Smallest.ai is not trying to be a general foundation model company. It is focusing on voice-specific details that can break a call: diverse accents, dozens of languages and noisy environments.

Inc42’s supplied report adds product detail, saying Smallest.ai announced Voice 4.0, powered by an in-house architecture called Hydra, and that its platform includes Pulse STT Pro for speech-to-text and Lightning V3.1 for text-to-speech. That report says the models support 38 languages and include enterprise features such as speaker diarisation, emotion detection and code-switching.

For buyers, the value proposition is less about novelty and more about replacing the awkward parts of phone automation: delayed replies, flat delivery and rigid scripts. Kamath’s argument is that support companies should not have to build deep voice infrastructure themselves.

When asked why well-funded AI customer support companies would not build their own models, Kamath said becoming “extremely good at doing voice is a distraction from their core business.”

That line also explains where Smallest.ai wants to sit in the stack. XOOMAR analysis: if the company can make voice latency and conversational realism hard to replicate, it becomes a specialist supplier to AI support companies rather than just another chatbot vendor.

For related XOOMAR coverage on how companies evaluate AI agents inside operating workflows, see Models Take a Back Seat in Target's AI Moat Strategy. For the investor side of AI startup formation, see Teen Founders Crash Silicon Valley’s AI Startup Club.

RingCentral, Truecaller and a crowded voice AI field set the enterprise test

Smallest.ai already counts companies in the voice space among its customers, including RingCentral and Truecaller, according to TechCrunch. Kamath also named customer support companies such as Sierra and Decagon as potential customers.

The competitive field is not empty. TechCrunch says Smallest.ai competes with ElevenLabs, Cartesia and regional players such as Sarvam.

The distinction matters. Some voice AI companies apply synthetic speech to dubbing, podcasting or broader media workflows. Smallest.ai is narrowing its focus to live enterprise calls, where interruption handling and response timing can decide whether a caller keeps talking or hangs up.

Kamath put the ambition in unusually direct terms.

“We want our models to break the Turing test,” Kamath said. “You should speak to our model and not know it's AI or human. That's the sole focus of the company.”

XOOMAR analysis: that goal is powerful commercially, but it also raises a trust problem. If an AI phone agent becomes difficult to detect, disclosure stops being a UX footnote and becomes a core product, legal and brand decision.

The next proof is messy calls, not polished demos

The raise gives Smallest.ai more room to build, but the harder test is outside the demo environment. Real customer calls include accents, bad microphones, background noise, interruptions, anger, silence and questions that jump across topics.

The supplied material says Smallest.ai is already focused on accents, dozens of languages and noisy environments. Those are not edge cases in phone support. They are the job.

The trust layer will also matter. Human-like AI calls create direct questions around consent, call recording, spam controls and whether callers should always be told they’re speaking with software. The source material does not say how Smallest.ai handles disclosure policies or customer deployment rules.

That is the practical watch item after this $13 million round: whether Smallest.ai voice AI can turn its latency thesis into reliable enterprise calls that feel useful rather than uncanny. If it can, customer support companies may decide voice is better bought from specialists. If it can’t, callers will hear the machine immediately.

The Bottom Line

• Smallest.ai is targeting latency, one of the biggest barriers to making AI phone agents feel human.
• The $13 million Series A shows investor confidence in specialized voice models over simply scaling general-purpose LLMs.
• If successful, the technology could reshape customer-service calls by making AI agents harder to distinguish from human operators.