Founders & Vision

The Team Behind the Engine: Useful Sensors & Moonshine

Meet Pete Warden and Manjunath Kudlur — TensorFlow founding members who built Moonshine, the speech model that runs on anything. Their work powers VoxBar Lite.

March 2026 • 4 min read

The Team Behind the Engine: Useful Sensors & Moonshine

VoxBar Lite is the one model in the suite that doesn't need a GPU. No NVIDIA card, no CUDA, no special hardware. It runs on anything — old laptops, office PCs, machines that other AI models won't even load on. That's possible because of a small company called Useful Sensors and a speech model called Moonshine.

Who They Are

Useful Sensors was founded by Pete Warden (CEO) and Manjunath Kudlur (CTO). If those names sound familiar, it's because they were both founding members of the TensorFlow team at Google — the framework that helped democratise machine learning for millions of developers worldwide.

After TensorFlow, they could have gone anywhere. They chose to start a company focused on running AI models on the smallest, cheapest hardware possible. Not in the cloud. Not on expensive GPUs. On the devices people already own.

That's the philosophy behind Moonshine.

What They Built

Moonshine is a speech recognition model designed from the ground up for edge deployment — meaning it runs locally on the device, with no internet connection, no cloud processing, and minimal hardware requirements.

Where models like Whisper and Parakeet were trained to maximise accuracy (and then worry about efficiency later), Moonshine was built to be small first. The result is a model that:

  • Runs comfortably on a CPU — no GPU needed at all
  • Uses roughly 200MB of disk space (compared to 1-5GB for other models)
  • Loads in about a second
  • Uses true streaming — audio flows directly into the model with no temp files, no disk I/O
  • Fires events in real-time as it recognises words

It's not the most accurate model in the VoxBar suite. It doesn't have the contextual intelligence of Canary Qwen or the benchmark-leading precision of Parakeet TDT. But it runs on hardware that those models can't even start on, and that matters.

Why It Matters to Us

Every other engine in the VoxBar suite requires an NVIDIA GPU. That's a real barrier. Not everyone has a dedicated graphics card. Not everyone can afford one. Not everyone wants one.

Moonshine removes that barrier entirely. It's the engine that means we can honestly say "VoxBar works on any PC." Without Moonshine, that wouldn't be true.

Pete and Manjunath built Moonshine because they believe AI should be accessible to everyone, not just people with expensive hardware. That aligns with what we're trying to do with VoxBar — private, local transcription that works for everyone, not just power users.

The TensorFlow Connection

There's something fitting about former TensorFlow team members building a model designed for the edge. TensorFlow's original mission was to make machine learning accessible. TensorFlow Lite extended that to mobile and embedded devices. Moonshine is, in a way, the logical next step — taking speech AI to the absolute minimum viable hardware.

The engineering pedigree shows. Moonshine's streaming architecture is elegant: audio flows in, events fire out, and the model manages its own internal state without the developer needing to handle buffering, chunking, or temp files. It's the kind of clean API design you'd expect from people who spent years building developer tools at Google.

Attribution

Moonshine is released under the Apache 2.0 license — fully open source, no attribution required, commercially permissive.

VoxBar is an independent product and is not affiliated with, endorsed by, or sponsored by Useful Sensors. We're just glad they built something that lets everyone use voice transcription, regardless of their hardware.