A protective communications headset is a small acoustic system with an unusually crowded electronics bill. Grinalds Solutions has developed the electronics for one such product, working for a hearing protection and communications manufacturer whose earcups carry seven microphones apiece, a four board stack, and a GreenWaves GAP9 class audio processor. What follows is a description of the constraints that shape that class of design. The constraints are physical, and they arrive in a fixed order.

The Passive Shell Sets the Ceiling

Active cancellation extends the attenuation curve downward. It does not repair the curve. In a hearing protection product, the passive earcup is doing most of the work above roughly 1 kHz, and the active system is buying attenuation in the region below that where cup mass and seal compliance run out.

The number on the box comes from a specific test. The Environmental Protection Agency requires the Noise Reduction Rating under 40 CFR Part 211 Subpart B, computed from real ear attenuation at threshold measured per ANSI S3.19-1974. That method uses ten subjects, three paired open and occluded trials each, and nine test bands, with the mean attenuation and the standard deviation both feeding the single number. The experimenter fit procedure is generous. ANSI/ASA S12.6 Method B, the subject fit method, generally returns lower values that track field performance more closely. NIOSH recommends derating earmuff labels by 75 percent, and OSHA applies a flat 50 percent derate. An engineering team should design against the derated field number.

The practical consequence for the electronics is the seal. Clamping force controls the leak area between the cushion and the head, and a leak bypasses the cup for low and mid frequencies entirely. Doubling clamping force roughly halves the leak, which is a mechanical lever and a comfort penalty. On the electrical side, a leak changes the plant that the feedback loop is trying to control. The transfer function from driver to internal microphone shifts, loop gain margins move, and a controller tuned on a sealed head and torso simulator can go unstable on a real head with glasses or a helmet strap under the cushion. Design the loop for the worst realistic seal, then verify across a fit population.

Microphone Placement, Count, and Matching

A hybrid topology needs at least two acoustic observation points per ear. Feedforward microphones sit outside the cup and see the disturbance before it arrives, which buys propagation time and allows useful cancellation up toward 1 kHz to 2 kHz. Feedforward paths do not correct for fit, and they are exposed to wind noise and handling noise. A feedback microphone sits inside the cup near the driver and observes the residual the wearer actually hears, which makes it tolerant of fit variation and of angle of arrival. Feedback paths are limited by cavity resonance and by loop stability, so gain has to be held back. Combining them gives independent control of the two paths and better coverage than either one alone, at the cost of processing load and microphone budget.

Situational awareness and voice pickup add more elements. A beamforming array steers sensitivity by exploiting the amplitude and phase relationship of the same wavefront across several spaced elements. That relationship only holds if the elements match. Part to part sensitivity spread and phase spread set the achievable null depth, and both matter more than absolute sensitivity. Measured on a real array, inter element delay shows up as a phase difference growing linearly with frequency, on the order of tens of degrees at 10 kHz for millimeter scale spacing. Any sensitivity or phase error adds directly to that budget.

Two microphone specifications drive the array choice. Signal to noise ratio sets the floor for quiet environments and for the feedback path, where microphone self noise becomes audible under cancellation. Acoustic overload point sets the ceiling. A protective headset lives in high level environments, and a feedforward microphone that clips on an impulse produces an anti-noise signal that is worse than doing nothing. Parts with an acoustic overload point in the 130 dB SPL to 146 dB SPL range with 65 dB to 80 dB signal to noise ratio are the relevant class. PDM output parts simplify the interconnect and keep analog runs short, which matters when the microphone is on a different board than the processor.

Latency Is the Real Budget

Cancellation is a phase problem. Every microsecond of group delay in the electrical path is a phase error that grows with frequency, and past the point where the error reaches 90 degrees the system adds energy rather than removing it. Acoustic propagation delay inside a cup is on the order of 5 to 10 microseconds and can generally be neglected. The digital path cannot. A workable target for the electrical group delay in a feedback loop is below about 5 microseconds, and published guidance for ANC signal chains puts the ideal end to end figure under 50 microseconds.

This is why conventional oversampled audio converters and frame based DSP do not work here. A codec with a decimation filter and a block processing pipeline can easily consume a millisecond, which restricts useful cancellation to the lowest octaves. Analog ANC avoids the problem by having almost no delay, and it remains a reasonable answer for a fixed filter feedback loop. Analog gives up adaptation, and it gives up the phase accuracy needed above roughly 1.2 kHz where the pressure field in the cup becomes position dependent.

The current generation of low power audio processors resolves this with dedicated filtering hardware rather than with faster general purpose cores. The GAP9 is representative. It carries ten RISC-V cores, an NE16 neural engine for inference, and a Smart Filtering Unit intended for PDM to PDM filtering at about 1 microsecond of latency, with three serial audio interfaces handling up to 48 audio channels in or out. The part ships in a 3.7 by 3.7 mm wafer level chip scale package. That combination is what allows a single device in one earcup to close a low latency hybrid ANC loop, run a beamformer, and still have compute left for neural network based noise reduction and scene classification.

Packaging the Electronics Into the Cup

Everything above has to fit behind a driver, inside a shell that also has to pass a drop test and a seal. In practice that means a stacked board set per cup rather than one board: a baffle board carrying the feedforward and feedback microphones at defined acoustic positions, a processor board, a power board, and an interface board. The boards are joined by flex, and the two cups are joined by a cable carrying a TDM link and power.

Three items on that interconnect deserve attention early. First, the serial audio link between cups is a clocked digital bus running through a flexible cable with limited impedance control, so source termination and edge rate need to be planned rather than discovered. Second, startup ordering matters. A transmitter that comes up before its receiver has a clock domain can leave a link silent in a way that looks intermittent. Third, a wafer level chip scale package with fine pitch is an assembly constraint, and confirming that the chosen assembler will place and inspect it should happen before the layout is committed.

None of this is exotic. It is a sequence: fix the passive attenuation and the seal, choose microphones for matching and overload rather than for headline sensitivity, hold the group delay budget, and then lay out boards that survive the mechanical package. Reordering those steps is what produces headsets that measure well on a fixture and disappoint on a head.

References