Neural audio for the software-defined vehicle.

For a century, the loudest thing in a car was the engine — and the engine, inconveniently, was also the best noise mask ever installed in a cabin. Electric vehicles removed it. What is left is quieter on average and harder in every other way: road texture, wind, rain, a conversation in the back seat. Noise that never repeats itself.

The dominant noise in a modern cabin is non-stationary — precisely the kind classical signal processing handles worst. At the same time the cabin gained more microphones, more speakers, more compute, and passengers who take calls, dictate messages, and expect the car to hear them the first time. The acoustic problem got harder exactly as the expectations got higher.

Hand-designed adaptive filters assume the world is roughly linear and slowly changing. Cabins are neither. Loudspeaker distortion is nonlinear, seating positions move, surfaces change with every stretch of asphalt, and the interesting events — a truck passing, rain starting — are transient by definition.

The industry’s answer has been calibration. Every vehicle program ends with weeks of manual tuning by acoustic engineers, repeated per trim, per market, per model year. It works — our team has done it for a decade — but it does not scale with the pace software-defined vehicles are asking for. The cost of classical audio is not the algorithm; it is the tuning.

Our chain has three blocks. A thin DSP front-end runs at sample rate: framing, transforms, linear echo cancellation. A neural core runs on the NPU at frame rate: suppression, separation, prediction, coefficient estimation. A thin DSP back-end returns to sample rate: cancellation filter execution behind a safety limiter.

Where a millisecond matters, the system stays deterministic. Where judgment matters, it is learned. In road-noise cancellation, the network aims and the DSP fires: learned models predict and adapt the controller, while cancellation itself executes at sample rate.

Audio is a real-time discipline. Acoustic paths do not wait for a scheduler, and an anti-noise wave that arrives late is a noise wave. So the parts of the chain where latency is physics — filter execution, limiting, the output path — run at sample rate in deterministic code, on a fixed budget, every block, without exception.

The limiter deserves its own sentence: it is a hard stage that no model output can bypass. If a learned component ever produces a pathological frame, the limiter bounds what reaches the loudspeakers and the passengers. Determinism here is not a limitation we accepted; it is the design principle that makes a learned system acceptable in a vehicle.

This split is also what makes the system reviewable. An OEM safety team can reason about the deterministic envelope — budgets, bounds, failure modes — independently of how the neural core was trained. The architecture is designed to functional-safety practice from the first block diagram.

Classical audio spends its money at the end of every program: design the filter, put it in a prototype vehicle, and tune by hand — for weeks — then do it again for the next trim and the next market. Learned audio moves that effort to data: record the cabin, train against it, fine-tune per platform, and keep improving after the vehicle ships.

The economics change shape. Manual calibration is a cost that repeats with every program; a data pipeline is a cost that amortizes across them. We are careful with claims here — our learned road-noise models are still ahead of us, and this site will carry measured numbers only when we have them — but the structure of the work is already different, and that is the part that compounds.

A learned audio product is not a checkpoint file. It is a model that must be executed on the vehicle’s NPU inside a strict budget, updated over the air without breaking the acoustic contract, and managed across silicon generations for the life of the car.

That execution layer is AREN, our automotive runtime. Audio is the first workload it carries; the layer itself is built to carry more. The AREN runtime page describes it at the same level of honesty as this one.

Our team has built production automotive audio for ten years, shipping software into OEM programs across six countries — echo cancellation, noise management, and sound systems that run in vehicles on the road today. The neural audio line is that experience, restated for what cabins have become.