Treble raises $18M to give physical AI a sense of hearing
Physical AI has been built on eyes and hands. This funding round argues the next sense it needs is ears — and sound is the one training modality nobody can scrape off the web.
Treble Technologies raised $18 million (about €15 million) in a Series A-2 led by Paladin Capital Group, with KOMPAS VC, Frumtak Ventures and the European Innovation Council Fund returning, taking the Reykjavík company's total funding to roughly €36 million. The pitch is deliberately unglamorous: physics-accurate acoustic simulation. Treble models how sound actually moves through a space — the geometry, the materials on the walls, where the speaker and the microphones sit — then turns those models into digital twins that generate labelled synthetic audio, so a device maker or a model developer can test a product before anyone books a recording studio.
That matters because audio has been physical AI's afterthought. Development budgets went to vision and touch; voice shipped as a feature and was evaluated in quiet rooms. Paladin's François Ruether framed the gap in the announcement: today's robots "may detect what is in front of them but cannot interpret critical sounds — such as someone falling in another room or a crash around a corner." The harder version is the moving one. A head-worn device, a humanoid, or a car changes its acoustic environment every second, and a model that scores well on clean speech can fall apart on reverberation, background noise, and a phone-grade microphone. Treble's claim is coverage: test cases that are expensive to stage even once become datasets, and a recording program that takes months becomes days.
The customer list is the proof, and it is not a startup's pilot list. Amazon and Logitech are both named in the round, and Amazon's senior audio research manager says Alexa development already leans on virtual acoustic environments to evaluate audio quality and scenarios "difficult to reproduce through physical testing alone" — the kind of endorsement a simulation vendor cannot buy. Earlier this year Treble partnered with Hugging Face on a benchmark for speech recognition models across realistic conditions, which is the same argument delivered to the people who rank models for a living. Paladin itself is a dual-use investor with government and commercial customers, which is the quiet signal in the round: audio perception is not only a consumer-hardware problem.
Why the synthesis angle is the real bet: as Treble chief executive Finnur Pind told TechCrunch, "pretty much all sound-related AI has been made from recordings and data scraped from the internet," and audio AI is fundamentally a data problem. If physics simulation can replace a meaningful share of that recording, the cost curve for voice and audio models bends the way image data bent a decade ago — and the companies that own the simulator, not just the datasets, sit upstream of the humanoid, wearable and automotive programs all racing to hear.
What to watch: whether the next round is led by defence-adjacent demand rather than consumer hardware, and whether any humanoid maker licenses the simulator itself instead of buying the generated data.
If a robot's hearing is simulated before it is built, does that count as experience — or just unusually good data? Tell us in the comments.
