Treble, an Icelandic startup that simulates how sound behaves in physical space, has raised $18 million in a round led by Paladin Capital Group. That takes total funding past $40 million for a company founded in 2020 by two acoustic engineers, Finnur Pind and Jesper Pedersen, which sells synthetic audio data and virtual device testing to the labs building voice models and to the hardware makers building the things that listen. Amazon and Logitech are among its customers. Existing investors KOMPAS VC, Frumtak Ventures, EIC and Omega ehf also took part.
The round is a step up from the $12 million Treble raised in 2024, and it arrives in the middle of a voice AI funding wave — billions going into customer support and sales automation, meeting note-takers, and smart glasses where speech is the primary control surface. Labs are shipping models quickly, device makers are chasing a comfortable interaction, and all of it has to be tested against something. Treble is not building any of those products. It is selling the test rig.
The pitch rests on a claim Pind makes about the foundations of the field: audio AI is first of all a data problem, and until now nearly every AI system that works with sound has been built from recordings and data gathered off the internet. Treble's position is that accurate physical simulation is an alternative way to produce audio data — you compute what a room does to a voice instead of going out and recording it.
That is the same substitution that already played out in autonomous driving, where simulated miles became a standard stand-in for road miles once the road miles stopped producing new edge cases. Acoustics has the same shape of problem. The interesting cases are rare, and they are rare because rooms, distances, materials and background noise combine in ways that no collection of recordings covers evenly.
For companies building voice AI, Treble sells a synthetic data generation platform whose output feeds speech enhancement, noise suppression and model training. It also evaluates voice models under varied conditions and passes the results back to the labs for refinement. Earlier this year it launched a benchmark with Hugging Face for speech recognition models under realistic conditions.
The second line of business is hardware. Treble works with headphone and speaker manufacturers on virtual prototyping, so a device can be heard before it is built. The platform can also test how well a smart speaker picks up commands depending on where it is placed. More recently the company has started running simulation tests for smart glasses and other AI devices.
Pind says Treble is interested in wearables that improve a user's hearing — a next generation of earbuds and smart glasses with something close to superhuman hearing. Such devices would help in difficult acoustic environments: in a restaurant, a user could keep audible only the people within two meters; at a seminar, they could mute the conversations around them.
The company now plans to push further into physical AI, working with firms building robots, cars and drones on sound-based features they can test and simulate. François Rüter, vice president at Paladin Capital Group, said Treble's platform stands out for its ability to simulate the behavior of different models and devices, and that its importance will grow as the company moves into new areas. Paladin's thesis is that acoustic infrastructure becomes more valuable as more products depend on understanding audio signals — voice AI, wearables, robotics, physical AI. Customers keep control of their own models, products and development workflows; Treble provides shared acoustic infrastructure that was built on simulation from the start.
The superhuman hearing scenario is the part of this announcement that will get quoted, and it is the part that matters least. It describes a product Treble does not sell, for a device its customers have not shipped. The physical AI line is the more consequential move, for a duller reason: to a robot, a car or a drone, a sound is a sensor reading, and the companies building those already buy simulation for vision and motion. Adding acoustics to an existing simulation budget is an easier sale than persuading a consumer electronics firm to change how it tests hardware it has been testing the same way for decades.
The number is worth reading carefully too. Just over $40 million across six years is modest beside the sums moving through voice AI applications, and deliberately so — Treble has positioned itself not to compete with the people it sells to. That is a defensible place to sit, but it caps how big the company gets on the growth of everyone else's audio ambitions.
What the announcement does not say is how closely simulated acoustics predict what actually happens in a room. That is the entire product. Nothing here reports what the Hugging Face benchmark found, how a Treble-simulated device compared with the built one, or by how much a synthetically trained noise suppressor beat a conventionally trained one. Amazon and Logitech as named customers is the only evidence offered, and a customer logo is a purchasing decision, not a measurement.
Every simulation business eventually faces the same audit: somebody builds the device, ships it, and compares what happens in the room with what the model said would happen. Treble now has $18 million and a widening list of things it has to be right about — earbuds, speakers, smart glasses, and soon robots, cars and drones — before that comparison is made in public.