Treble Technologies has raised $18 million to expand an acoustic-simulation business that is moving beyond building design and audio engineering into the training and testing of AI systems expected to understand the physical world through sound.

The Reykjavík company’s Series A-2 round was led by Paladin Capital Group, with existing investors KOMPAS VC, Frumtak Ventures, and the European Innovation Council Fund also participating. Treble says the latest financing takes its total funding to €36 million and will support further US expansion alongside product development for audio-enabled devices and physical AI.

Treble’s technology models the way sound travels through rooms, vehicles, devices, and other environments, allowing developers to create digital acoustic twins and generate labelled synthetic audio rather than record every combination of voice, noise, room geometry, microphone position, and material in the physical world. Customers cited by the company include Amazon and Logitech.

The expansion reflects a practical problem emerging as AI moves into wearables, robots, vehicles, and other machines. Computer vision receives much of the development attention, but products expected to hear speech or recognise events also have to function in acoustically messy environments where reverberation, multiple speakers, background machinery, and changing device positions materially alter what a microphone captures.

Synthetic data moves into the acoustic world

Training speech and sound models traditionally requires large volumes of recordings, and collecting representative data can become expensive once developers try to account for different rooms, speakers, surfaces, microphones, distances, and sources of background noise. Some edge cases may be uncommon enough that gathering them repeatedly under controlled conditions is impractical, while hardware that has not yet been built cannot be tested physically at all.

Simulation offers another route by allowing a development team to alter the environment computationally and generate consistent datasets around the resulting acoustic conditions. Treble says its platform can represent different rooms, materials, vehicles, device configurations, speakers, and ambient sounds, while its digital-twin approach allows hardware and AI behaviour to be evaluated before a full physical prototype or recording programme exists.

That can make simulation useful before model training as well as during it. A device maker deciding where microphones should sit can compare configurations virtually, while an AI team can test whether a speech-recognition model becomes unreliable when a person moves away from the device, another speaker talks nearby, or surrounding surfaces introduce stronger reverberation.

Treble says the approach can reduce some prototype testing and data-collection programmes from months to days, although that figure comes from the company and will depend heavily on the product and validation requirements involved. Physical testing does not disappear because simulation becomes faster; instead, simulation can narrow the configurations that need to reach expensive real-world testing and expose failure modes earlier in development.

Simulation is useful only when reality agrees

The central technical risk around synthetic training data is whether the modelled environment reproduces the characteristics that matter once a device leaves the lab. An acoustic simulator can generate almost unlimited variations, but volume alone does not compensate for inaccuracies in the physics, assumptions, microphone models, or environmental information underneath those samples.

Developers consequently need to validate simulated outputs against measured real-world data rather than treating synthetic information as an automatic replacement. That resembles the challenge already visible elsewhere in AI simulation: a model trained successfully in a digital environment can still encounter a gap when sensors, objects, people, weather, noise, or unpredictable behaviour differ from the assumptions represented during training.

Acoustics adds its own difficulties because minor physical changes can alter the signal reaching a microphone. Room dimensions, wall materials, furnishing, movement, microphone direction, device casing, speaker position, and competing noise all influence the final recording, while a wearable or mobile robot changes its acoustic relationship with the environment as it moves.

Treble’s commercial opportunity therefore rests less on the ability to generate sound files than on whether its simulation remains sufficiently accurate to support engineering decisions. The company describes its underlying modelling as physics-accurate and has historically developed wave-based acoustic simulation technology, a computationally demanding approach intended to reproduce phenomena that simpler methods may approximate less precisely.

AI products are becoming sensor systems

The move into physical AI also illustrates how AI development is broadening beyond text prompts and software workflows. A robot, smart headset, industrial machine, or vehicle may need to combine cameras, microphones, motion sensors, spatial information, and other inputs before it can determine what is happening around it.

Sound can provide information that vision cannot. Equipment faults may be audible before they are visually apparent, speech can come from outside a camera’s field of view, and an impact, alarm, or other event may occur behind an obstruction. Combining modalities potentially produces a richer understanding of the environment, although it also increases the amount of data that has to be generated, labelled, processed, and validated.

For manufacturers, that expands AI development into a hardware-engineering problem as much as a model problem. Microphone selection, placement, enclosure design, onboard compute, power use, connectivity, latency, and privacy controls all affect whether an audio model can function reliably in the finished product.

Treble already sells simulation tooling into engineering and product-development workflows, giving it a route into organisations before a model reaches deployment. Its challenge after the financing round will be demonstrating that the same modelling technology can become part of the data infrastructure behind AI products rather than remaining primarily a specialist acoustics tool.

As AI systems move away from clean digital inputs and into rooms, streets, vehicles, factories, and devices worn on the body, developers face a growing collection problem for every sensor modality they add. Treble is betting that sound does not need to be recorded in every possible environment before machines can learn how to hear it, provided the simulated version is close enough to the world in which those machines eventually operate.