From hand-tuned to data-trained

title: "From hand-tuned to data-trained" subtitle: "Why in-cabin audio is the first automotive domain the learned stack should win — and what it takes to run it on the car." author: "Arenn.ai" date: "2026-08-03" tags: [neural-audio, essay]

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. It was broadband, it was predictable, and it rose exactly when the other noises rose: accelerate, and the engine covered the road; cruise, and it covered the wind. Generations of acoustic engineers learned to work with it the way stage designers work with lighting. You did not need to remove every unwanted sound. You needed the dominant one to be the one you controlled.

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.

A short history of quiet

Automotive audio has always been a discipline of the possible. The cabin is one of the worst acoustic environments in consumer technology: a small reflective box with glass on five sides, seats that absorb unevenly, loudspeakers mounted wherever the door structure allowed, and microphones placed by an interior designer rather than an acoustician. Into this box the industry has shipped, over decades, a remarkable stack of signal processing — echo cancellers so a phone call survives the school run, equalization tuned per trim level, engine-order cancellation phased against the crankshaft, and lately road-noise systems that push anti-noise through the same speakers that carry the music.

Nearly all of it was built the same way. An algorithm — usually an adaptive filter with well-understood mathematics — was designed offline, then brought to a prototype vehicle, and then tuned. The tuning is the part outsiders underestimate. It is weeks of trained engineers in parked and moving cars, adjusting coefficients, re-measuring, listening, and adjusting again, for every vehicle variant, every seat configuration that matters, and every market with different regulatory and taste expectations. The algorithm was general; the quiet was handmade.

The economics of that loop were tolerable for as long as two things held. Vehicle programs moved slowly enough that a tuning season fit inside them, and the dominant noise source — the engine — was stationary enough that a filter tuned in March still held in November. Both assumptions are now failing at once.

What electric vehicles actually changed

The obvious change is the missing engine. The subtler change is statistical. Engine noise is, in signal-processing terms, a friendly opponent: strongly periodic, correlated with measurable quantities like RPM, and slow to change its character. A cancellation system can be told, by the crankshaft itself, what the enemy will sing next.

Road noise has no crankshaft. It is the sum of tire tread against changing asphalt, cavity resonances, suspension transmission, and aerodynamic turbulence — a signal that is broadband, non-stationary, and only weakly predictable from any single sensor. Rain arrives without an order number. A passing truck is a transient. A conversation in the back seat is, from the front microphones' point of view, both signal and noise depending on who is on the phone.

The statistical character of the problem moved from periodic to chaotic exactly as the masking floor dropped. Passengers hear more of everything, and what they hear is harder to model. Meanwhile the cabin itself became a computing environment: more microphones for voice assistants, more loudspeakers for spatial audio, and — this is the part that matters for this essay — a neural processing unit sitting on the vehicle's main SoC, mostly idle, paid for by the driver-assistance budget.

Why the classical toolbox plateaus

Classical adaptive filtering rests on assumptions that cabins now violate daily. The world is assumed roughly linear; loudspeakers driven hard are not. The environment is assumed slowly varying; a wet on-ramp is not. The reference sensors are assumed to observe the noise before it reaches the ear; for many road-noise paths, the acoustic head start is measured in single-digit milliseconds, and the filter must commit to its output long before the statistics settle.

Engineers have answered each violation with structure: more reference accelerometers, band-split processing, gain schedules keyed to vehicle speed, hand-built detectors for rain and rough asphalt. Each answer works, and each answer adds another dial that a human must set per vehicle, per trim, per market. The tuning season grows with the ambition.

This is the plateau. It is not that classical methods stopped working — they ship today, including ours. It is that their remaining headroom is purchased almost entirely with human calibration time, and calibration time is precisely what software-defined vehicle programs no longer have. The cost of classical audio is not the algorithm; it is the tuning.

What learning changes — and what it does not

A learned model changes where the knowledge lives. Instead of a human encoding the cabin's behavior into filter structures and schedule tables, the behavior is recorded — microphones, accelerometers, driving conditions — and a model is trained against it. The knowledge moves from the engineer's calendar into the dataset.

Three properties of the audio problem make this more than a fashionable substitution. First, supervision is cheap: for suppression and enhancement tasks, the ground truth is often obtainable by construction — mix the noise in, ask the model to take it out — so the data pipeline does not depend on armies of human labelers. Second, the domain is mercifully low-dimensional compared to vision: a cabin's acoustic state fits in kilobytes per frame, so models sized for a vehicle NPU are not a compromise but a natural fit. Third, evaluation is honest: unlike open-world driving, an audio system's failures are audible, measurable, and reproducible in a seat, which keeps the feedback loop short.

None of this makes the network magic. A learned suppressor trained on the wrong distribution fails like any other software, only less legibly. The serious work is unglamorous: which cabins, which surfaces, which seat states, which languages of conversation, and how the model behaves on the day it meets a combination it never saw. That work is engineering, not alchemy, and it is exactly the kind of work a team raised on production tuning seasons knows how to schedule.

Deterministic where it must be

Audio is a real-time discipline, and no amount of learning changes the physics. An anti-noise wave that arrives late is a noise wave. So the architecture we build is deliberately hybrid, and unapologetically so.

A thin DSP front-end runs at sample rate: framing, transforms, linear echo cancellation — the operations where microseconds are the budget and the mathematics is settled. A neural core runs on the NPU at frame rate: suppression, separation, prediction, coefficient estimation — the judgments where the world is messy and learning pays. A thin DSP back-end returns to sample rate: the cancellation filters execute, and a hard limiter bounds everything that reaches the loudspeakers.

Where a millisecond matters, the system stays deterministic. Where judgment matters, it is learned. In road-noise cancellation the division of labor is exact: the network aims, the DSP fires. Learned models predict the noise environment and adapt the controller; cancellation itself executes at sample rate in deterministic code.

The limiter deserves its own paragraph, because it is the sentence a safety reviewer reads twice. It is a fixed, inspectable stage that no model output can bypass. If a learned component ever produces a pathological frame, the limiter bounds what the passengers hear. Determinism in this architecture is not a concession to old habits. It is the property that makes a learned system reviewable — an OEM team can reason about budgets, bounds, and failure modes without needing to audit how the network was trained.

The economics, restated

Put the two loops side by side. The classical loop: design a filter, bring it to a prototype vehicle, tune by hand for weeks, then repeat — per trim, per market, per model year. The learned loop: collect cabin data, train and fine-tune, deploy on the vehicle, and keep improving over the air after the program ships.

The first loop spends its money at the end, every time. Its marginal cost per program is roughly constant, because the expensive ingredient — expert attention in a physical seat — cannot be amortized. The second loop spends its money in the middle, once, on data infrastructure and training discipline, and its marginal cost per program falls as the dataset and the tooling compound. A new trim becomes a fine-tune. A new market becomes a data slice. A field complaint becomes a training example rather than a service campaign.

We state this as a change of shape, not a promise of magnitude. The honest version of the claim is structural: hand-tuning cannot compound, and data can.

What is unsolved

An essay that pretends the hard parts are behind us would be marketing, so here are the hard parts in plain terms.

Generalization across cabins is not free. A model trained in one vehicle meets different transfer paths in the next, and the discipline of making adaptation cheap — what to share, what to fine-tune, what to measure before trusting it — is engineering we are still doing. The reference problem does not disappear because the filter is learned; microphones and accelerometers still observe the world late and incompletely, and prediction has limits that physics, not optimism, will set. On-device constraints are real: NPU schedulers are built for vision frame rates, audio needs its answers in tighter and more regular slices, and making a neural core keep a frame-rate promise on shared silicon is precisely the sort of dull, decisive problem that decides whether any of this ships. And evaluation must grow up with the system: a suppressor can score beautifully on averaged metrics and still produce the one artifact — a musical gurgle, a pumping floor — that a trained ear rejects in seconds. The metric that matters is a person in a seat.

We write these down because they are the work. A reader who has shipped vehicle software will recognize the list; a reader who has not should at least come away knowing that the difference between a demo and a program lives here.

The layer beneath

There is one more consequence, and it is the quiet one. A learned audio product is not a checkpoint file. It is a model that must execute on the vehicle's NPU inside a strict budget, be updated over the air without breaking an acoustic contract, coexist with other workloads on shared silicon, and remain manageable across hardware generations for the life of the car. Hand-tuned audio never needed that machinery; data-trained audio does not work without it. That execution layer — model runtime, update path, lifecycle — is a piece of infrastructure the software-defined vehicle will need for every learned workload, not only ours. Audio happens to be the first tenant. We build the layer as carefully as the tenant, and we suspect, without ceremony, that the layer is the larger of the two.