Smallest.ai, a startup that develops artificial intelligence models optimized to generate speech, has raised $8 million in seed funding.

Sierra Ventures led the investment. Smallest.ai, officially Smallest Inc., said in its announcement of the deal today that 3one4 Capital and Better Capital contributed as well.

Smallest.ai was founded last year by Chief Executive Officer Sudarshan Kamath and Chief Technology Officer Akshat Mandloi. The duo previously held engineering roles at Robert Bosch GmbH. The company’s initial focus is developing speech models for contact center teams, which can use the technology to automatically process customer requests.

Smallest.ai has trained an algorithm called Lightning that it claims is the fastest text-to-speech model on the market. It can generate 10 seconds of speech in 100 milliseconds, which corresponds to one-1oth of a second. That makes Lightning 50 times faster than certain competing models.

According to Smallest.ai, the algorithm’s performance stems from the fact that it uses a different design than many alternatives. Most voice models are based on an auto-regressive architecture that generates speech one token, or unit of data, at time. Lightning, in contrast, generates multiple tokens at once to save time.

The model requires less than a gigabyte of VRAM, the memory built into graphics processing units. Smallest.ai says it achieved that compact memory footprint by removing unnecessary weights from Lightning and compressing the rest.

The company debuted its newest voice model, Electron v2, late last month. Its has a TTFT, or time to first token, of 53.25 milliseconds, which means it takes about 0.05 seconds to start generating a response to user prompts. That makes it useful for latency-sensitive use cases such as answering customer support requests.

Electron v2 features 4 billion parameters. Smallest.ai claims that it can match the output quality of algorithms six times its size.

According to the company, customers can customize its models by uploading an audio snippet and instructing them to replicate the speaker’s voice. Smallest.ai provides two customization options. The first enables its models to clone a voice based on a 15-second recording, while the other requires 15 to 45 minutes of audio and produces higher-quality output.

Smallest.ai also enables users to customize its algorithms in other ways. Companies can create industry-specific voice agents optimized for tasks such as processing credit card numbers. According to Smallest.ai, organizations in industries with particularly strict cybersecurity requirements have the option to deploy its software on-premises.

“Our in-house models aren’t just about speed — they’re about enterprise reliability,” Mandloi said. “We’ve engineered natural speech understanding, guardrails and expressiveness into a single coherent stack.”

The company will use its new funding to grow its market presence in the retail, healthcare and technology sectors.

Photo: Unsplash