Smallest.ai raises $13 million to build full-duplex enterprise voice AI
Seligman Ventures led the round as founders Sudarshan Kamath and Akshat Mandloi introduced the Hydra speech-to-speech architecture.
By Ryan Merket · Published
Why it matters
Voice-agent margins depend on latency and per-minute inference costs. Smallest.ai is financing a vertically integrated stack designed to control both.
Sudarshan Kamath and Akshat Mandloi have raised a $13 million Series A for Smallest.ai, financing their attempt to replace the chain of models behind enterprise voice agents with a system that can listen and respond in parallel.
Seligman Ventures led the financing, with Sierra Ventures and 3one4 Capital participating, Smallest.ai said in a July 30th announcement. The round brings Smallest.ai's total funding to more than $21 million, following an $8 million seed led by Sierra in October 2025.
Kamath and Mandloi are IIT Guwahati mechanical engineering graduates who moved into artificial intelligence before founding Smallest.ai. Both worked on autonomous-vehicle systems at Bosch. Kamath later led product work at legal-services platform Vakilsearch and earned a master's degree in data science from the University of California, San Diego. Mandloi's background includes computer-vision and data-pipeline work for autonomous systems.
Their founding thesis runs against the tendency to route every AI workload through a large general-purpose model. Smallest.ai argues that specialized models, connected to outside memory and tools, can run faster and closer to the user. Voice is where that efficiency becomes immediately perceptible: delays measured in fractions of a second can determine whether a conversation feels fluid or mechanical.
A bid to own the entire voice loop
Most generative voice agents pass a caller's audio through several components. Speech recognition produces text, a language model decides what to say, and a text-to-speech model generates the response. Orchestration software, memory and safety controls add more steps. Each handoff introduces another source of latency, cost and failure.
Smallest.ai calls its proposed replacement "Voice 4.0," a marketing label for an asynchronous architecture that processes listening, reasoning, actions and responses concurrently. The product at its center is Hydra, a full-duplex speech-to-speech model that receives microphone audio and streams generated audio back through a single WebSocket connection.
The technical distinction matters. Hydra handles audio directly rather than emitting a transcript stream, according to Smallest.ai's documentation. Applications that need text must separately transcribe the incoming or outgoing audio using Smallest.ai's Pulse speech-recognition model. The documentation also describes support for "barge-in," allowing an application to discard a generated response when a user interrupts and a new response begins.
Kamath's case for the architecture is grounded in how people converse. "Humans don't wait for someone to finish speaking before they begin thinking," he said in the funding announcement. Smallest.ai is trying to reproduce that overlap without forcing enterprises to assemble speech recognition, reasoning and speech generation from separate suppliers.
Hydra joins a product line that includes Pulse STT Pro and Lightning V3.1, Smallest.ai's text-to-speech model. Smallest.ai says Pulse supports 38 languages, speaker identification, emotion detection, code-switching, noise reduction and redaction of personally identifiable and payment-card information.
Independent benchmark data provides a narrower view of performance than the funding announcement. Artificial Analysis listed the July 2026 release of Lightning V3.1 Pro eighth on its text-to-speech leaderboard, priced at $19.50 per million characters. That placement supports Smallest.ai's cost argument while stopping short of establishing it as the highest-quality model in the field.
Enterprise economics drive the architecture
Smallest.ai is targeting contact centers, financial services, healthcare providers and business-process outsourcers, where voice systems must handle interruptions, background noise, sensitive data and large call volumes. Smallest.ai names RingCentral, Truecaller, Readymode, Piramal, Kogta and Pocket among the organizations using its technology. Those customer relationships and their scale have not been independently quantified.
Smallest.ai also claims customers have reduced support costs by as much as 80% and increased agent productivity by up to 10 times. The announcement does not provide customer-level data, measurement periods or baselines for those figures, so they remain Smallest.ai's performance claims.
The architecture is tied directly to the economics of selling calls by the minute. Smallest.ai's published agent pricing ranges from about $0.09 to $0.21 per minute depending on the selected models and pipeline. The page lists a $0.01-per-minute hosting fee, additional model and telephony costs, and phone-number rental at $10 per month. Enterprise contracts use custom pricing.
Owning speech recognition, speech generation, orchestration and speech-to-speech inference gives Smallest.ai more control over those unit costs. It also creates a heavier technical burden: Smallest.ai must keep each layer competitive against specialized vendors while supporting the reliability and monitoring requirements of enterprise call operations.
Deployment is another part of that sales pitch. Smallest.ai offers on-premises and on-device inference, which can reduce network round trips and keep data inside a customer's infrastructure. Smallest.ai says its enterprise offering supports role-based access, customer-controlled logs and tracing, and compliance with SOC 2 Type 2, ISO 27001, HIPAA and GDPR requirements.
An investor follows the founders across firms
The Series A maintains continuity with Smallest.ai's seed round. Ashish Kakran, now a managing partner at Seligman Ventures, publicly championed Sierra Ventures' $8 million seed investment while he was a partner at Sierra. He is now backing Kamath and Mandloi from Seligman, which launched a $500 million venture arm in February 2026 focused on AI infrastructure, cloud infrastructure, cybersecurity and data-center technology.
Sierra and 3one4 also returned for the Series A. That gives Smallest.ai a lead investor already familiar with its architecture and enterprise case, while the previous institutional backers continue financing the buildout.
Smallest.ai reported nearly 60 employees in the July 30th announcement and said it plans to expand over the next year. The new capital arrives less than 10 months after the seed round, compressing the period in which Kamath and Mandloi must turn their technical thesis into repeatable enterprise deployments.
Voice agents have progressed far enough that sounding convincing in a controlled demonstration is no longer the main test. Production systems are judged on interruption handling, uptime, compliance, latency and cost over millions of minutes. Kamath and Mandloi are betting that controlling the full stack gives Smallest.ai enough room to improve all five at once.