Smallest.ai

Smallest.ai builds small, specialised speech models — text-to-speech, speech-to-text, speech-to-speech and a compact language model — plus a platform for deploying real-time voice agents. It targets enterprises in regulated sectors such as financial services, healthcare and contact centres that need interruptible, low-latency phone conversations rather than narration or dubbing, and it competes on speed and cost per minute rather than expressive range.

Smallest.ai is a voice AI company built on the thesis that real-time conversation needs many small specialised models rather than one large general-purpose one. Its stack is four models plus a deployment platform: Lightning, a text-to-speech model quoting sub-100ms time-to-first-audio across 15 or more languages at 44.1kHz; Pulse, speech-to-text covering 38 languages with speaker and emotion detection; Electron, a roughly four-billion-parameter small language model that handles the real-time conversational turn and hands off to a larger offline LLM only when a query exceeds its knowledge; and Hydra, a native speech-to-speech model whose asynchronous architecture runs listening, reasoning, acting and responding in parallel to cut the dead air that makes an interruptible phone call feel wrong. These sit behind Atoms, a dashboard, playground and API for configuring voices, languages and agent behaviour and deploying to rented phone numbers or a web widget. Founded in 2023 by Sudarshan Kamath and Akshat Mandloi, the company employs roughly 60 people and closed a $13 million Series A on 30 July 2026 led by Seligman Ventures with Sierra Ventures and 3one4 Capital participating, taking total funding past $21 million within about nine months of an $8 million seed. It launched Voice 4.0 alongside the round and reports a 76% naturalness win rate against OpenAI's gpt-4o-mini-tts, a 5.38% word error rate and a 3.89 mean opinion score. RingCentral and Truecaller are named customers. The company sells into financial services, healthcare and contact centres with SOC 2 Type 2, ISO 27001, GDPR, HIPAA and PCI compliance plus on-premise and private-cloud deployment, and frames its edge as hardware-level: roughly 550 simultaneous voice calls on about $27,000 of P100 accelerators against around $100,000 of NVIDIA L40S GPUs.

Category

Audio & Voice

Pricing

Usage-based, Freemium, Contact for pricing

Target Market

CTOs, VP Customer Experience, Contact Centre Leaders, Enterprise Developers, Product Engineers

Deployment

Cloud-first, API-based, Self-hosted, Hybrid

Founded

2023

Team Size

51-200

• ✓
  Lightning text-to-speech

  Sub-100ms time-to-first-audio at 44.1kHz across 15 or more languages, with prosody tuned for conversational turn-taking rather than narration.

• ✓
  Hydra speech-to-speech

  Asynchronous architecture that listens, reasons, acts and responds in parallel, cutting the lag that makes interruptions feel unnatural on a call.

• ✓
  Pulse speech-to-text

  Transcription across 38 languages with speaker and emotion detection, so an agent can react to tone as well as to words.

• ✓
  Electron small language model

  A roughly four-billion-parameter model handling the real-time turn, escalating to a larger offline LLM only for genuinely complex queries.

• ✓
  Atoms agent platform

  Dashboard, playground and API for configuring voices, languages and behaviour, then deploying to phone numbers or a web widget.

• ✓
  Fast voice cloning

  Production voice cloning from as little as five seconds of reference audio, avoiding the studio recording sessions rivals require.

• ✓
  On-premise and private-cloud deployment

  The full model stack runs inside the customer's own environment, which is what makes regulated healthcare and banking deployments possible at all.

• •
  Contact-centre call deflection

  A bank routes routine balance and payment queries to a voice agent that callers can interrupt without the conversation falling apart.

• •
  Outbound appointment and collections calling

  Healthcare or lending teams run high-volume outbound campaigns where per-minute model cost decides whether the unit economics work.

• •
  Multilingual customer support

  A support line handles callers who code-switch mid-sentence, with the agent following the language change rather than losing the thread.

• •
  Embedding speech into an existing product

  A software vendor adds voice input and output to its own application through the API without building a speech stack in-house.

• •
  On-premise voice for regulated data

  A hospital deploys the models inside its own infrastructure so patient audio never crosses the network boundary to a third party.

Metered per minute of conversation rather than per seat: $0.09 to $0.21 all-in depending on model choice, and unusually the vendor publishes the component breakdown — roughly $0.009 per minute for speech-to-text, $0.09 for text-to-speech, $0.01 flat for hosting, plus whichever LLM you route to at $0.045 to $0.10 per minute. Phone numbers rent at $10 per month each and text messages cost $0.005. New accounts get $10 in free credits. Everything that makes it deployable inside a bank or hospital — on-premise hosting, SSO, the 99.99% SLA, dedicated infrastructure — is Enterprise-only and unpriced, and the self-serve tier caps concurrency at 20 calls, so real contact-centre volume forces a sales conversation.

This page was written from 5 sources, 3 on domains other than smallest.ai.

1. 1.smallest.ai — smallest.aivendor
2. 2.smallest.ai — pricingvendor
3. 3.techcrunch.com — smallest ai raises 13m to build ultra fast voice ai that sou
4. 4.cmswire.com — smallestai lands 13m series a for voice ai
5. 5.cartesia.ai — elevenlabs vs smallest