Ex-Googlers take AI offline for billions of devices across Asia

“Talking to smart-home devices is like trying to talk to somebody who’s hard of hearing,” Pete Warden recalls someone telling him. “You know it only responds to specific phrases, but you don’t know what those phrases are.”

Warden is the co-founder and CEO of Moonshine AI. The US-based startup develops voice translation and speech processing software designed to run offline on everyday devices – like kitchen appliances or speakers – without sending data to the cloud.

A former Google engineer for machine learning platform TensorFlow, Warden thinks the current experience with IoT devices is subpar. These tools are often built to expand companies’ revenue streams rather than actually helping people interface better with electronics through voice, he says.

To solve that problem, he built his own company.

Image credit: Timmy Loen

“I want to be able to buy a lamp, plug it in, say ‘on,’ and have the lamp come on. I believe that’s how consumers want things to work,” Warden tells Tech in Asia.

Investors like Amplify Partners, Wing VC, and Ascend Vietnam Ventures quickly came on board. Moonshine raised US$9 million in 2022 and US$3 million in July 2025. It expects to raise a series A round in 2026.

Having secured government and enterprise contracts in the US and Australia, Moonshine is now eyeing Asia as another key market. The region is driven by the huge scale of local electronics manufacturing, and current AI speech models are “not usable enough” for Asian languages, Warden notes.

Tiny builds, bigger impact

Together with his fellow TensorFlow team member Manjunath Kudlur, Warden co-founded Moonshine in March 2022 under the name Useful Sensors. The initial idea was to develop small hardware modules that can be embedded into various devices to add features like person detection and voice control.

However, the team realized that this approach meant it may have a slow product lifecycle, as they would need to work with manufacturers from the early design stage.

Moonshine aims to integrate its voice AI models into everyday electronics and appliances / Photo credit: Lazy_Bear / Shutterstock

“It might be another three years until that comes to market. But by providing software, we can actually come in much later in the appliances lifecycle,” Warden says. So, the company switched its focus from hardware to software around 1.5 years ago.

The concept of local speech processing itself isn’t new. Transcription tools for meetings that work offline – such as Slipbox and Hyprnote – exist. These tools typically use lightweight, open-source AI models such as OpenAI’s Whisper or Nvidia’s Parakeet to process audio into text directly.

Challenging Big Tech

Smaller might not be better