Tag: FxLMS

Modern acoustic isolation in reference headphones and in-ear monitors relies on a sophisticated convergence of passive mechanical sealing and real-time digital signal processing (DSP). Active Noise Cancellation (ANC) operates on the fundamental physical principle of destructive wave interference: by sampling ambient environmental acoustic pressure via dedicated miniature electret or MEMS microphones, inverting the phase of the incoming acoustic waveform by 180 degrees ($\pi$ radians), and injecting that anti-noise signal into the headphone transducer’s acoustic chamber, low-frequency environmental noise is algebraically nulled before reaching the human tympanic membrane.

While the theoretical concept of destructive interference is straightforward, executing real-time wideband acoustic cancellation inside a complex, non-linear earcup cavity presents severe electro-acoustic control theory challenges. Acoustic propagation delays between the external sense microphone and the internal loudspeaker driver require ultra-low-latency analog-to-digital converters (ADCs) and Filtered-X Least Mean Squares (FxLMS) adaptive digital filters operating at sampling rates exceeding 384 kHz. This comprehensive engineering guide examines passive versus active attenuation spectrums, feedforward, feedback, and hybrid ANC topologies, secondary path transfer functions, occlusion effect mitigation, and transparency mode DSP architectures.

Sound propagates through air as longitudinal pressure waves consisting of alternating compressions and rarefactions. When two coherent acoustic waves occupy the same spatial volume simultaneously, the resulting instantaneous acoustic pressure $P_{total}(t)$ is governed by the principle of linear superposition:

Primary Noise Wave: P_noise(t) = A * sin(2*pi*f*t)
Anti-Noise Wave:    P_anti(t)  = (A + dA) * sin(2*pi*f*t + pi + dPhi)

Residual Error Pressure (dB Attenuation):
Attenuation (dB) = 20 * log10( sqrt( 1 + (1 + dA/A)^2 - 2*(1 + dA/A)*cos(dPhi) ) )

Key Engineering Constraint:
- To achieve 20 dB of active attenuation, phase error (dPhi) must be < 5.7 degrees, and amplitude error (dA/A) must be < 0.8 dB.
- To achieve 30 dB of active attenuation, phase error (dPhi) must be < 1.8 degrees, and amplitude error (dA/A) must be < 0.27 dB.

Because the wavelength $\lambda = c / f$ of sound decreases rapidly as frequency increases, maintaining sub-2-degree phase accuracy across a physical earcup cavity becomes exponentially more difficult above $1\text{ kHz}$. At $100\text{ Hz}$, the acoustic wavelength in air ($c = 343\text{ m/s}$) is $3.43\text{ meters}$ ($3,430\text{ mm}$). A $2\text{ degree}$ phase tolerance at $100\text{ Hz}$ corresponds to a time budget of $55.5\text{ microseconds}$ ($\mu\text{s}$) or a spatial tolerance of $19\text{ mm}$. Conversely, at $2,000\text{ Hz}$, the wavelength shrinks to $171.5\text{ mm}$, reducing a $2\text{ degree}$ phase window to just $2.77\text{ }\mu\text{s}$ ($0.95\text{ mm}$).

Consequently, active noise cancellation is overwhelmingly effective in the low-frequency spectrum ($20\text{ Hz}$ to $800\text{ Hz}$)—precisely where turbofan aircraft engines, subway rail rumble, and HVAC compressors concentrate their acoustic energy. Above $1\text{ kHz}$, passive mechanical isolation provided by high-density viscoelastic memory foam earpads, damped polycarbonate shells, and acoustic labyrinths takes over, absorbing short-wavelength high-frequency acoustic energy that would otherwise violate the phase latency limits of the electronic ANC loop.

In a Feedforward ANC topology, the reference sense microphone is mounted on the exterior surface of the headphone earcup or earbud shell, facing outward toward the ambient environment. The microphone captures the approaching environmental noise wave $x(n)$ before it penetrates the passive acoustic barrier of the earpad cushion.

For a feedforward system to cancel noise inside the ear canal, the total electronic processing time $\tau_{DSP}$ (comprising microphone transduction, preamplification, ADC conversion, digital filtering, DAC conversion, and driver actuation) must be strictly less than or equal to the physical acoustic propagation time $\tau_{acoustic}$ required for the external sound wave to travel from the outer microphone port through the earpad seal into the ear canal:

Causality Condition:
tau_DSP = tau_mic + tau_ADC + tau_filter + tau_DAC + tau_driver <= tau_acoustic

Typical Physical Values (Circum-aural Headphone):
- Distance from outer mic to ear canal: d ≈ 25 mm to 35 mm
- Acoustic Time-of-Flight (tau_acoustic): d / 343 m/s ≈ 73 us to 102 us
- Required Total Electronic Loop Latency: < 70 microseconds

Because the feedforward processor only observes the external noise $x(n)$ and does not have an internal microphone listening to the actual residual result inside the earcup, it is an open-loop control system. Its transfer function $W(z)$ must precisely model the ratio between the passive acoustic leakage path $P(z)$ (how external noise warps as it squeezes through the earpad foam) and the electro-acoustic secondary path $S(z)$ (how the headphone driver's output propagates into the ear canal):

$$W(z) = -\frac{P(z)}{S(z)}$$

While feedforward ANC cannot self-oscillate (making it unconditionally stable) and provides wide bandwidth cancellation up to $1.5\text{ kHz}$, it suffers from a critical vulnerability: fit sensitivity. If the user wears thick eyeglasses or shifts the headphone on their head, the passive acoustic leakage transfer function $P(z)$ changes dramatically. A static feedforward filter calibrated on a laboratory Head and Torso Simulator (HATS) will now inject an anti-noise signal with the wrong amplitude and phase, severely degrading cancellation or even inadvertently amplifying specific mid-band frequencies.

To overcome the blind open-loop limitation of feedforward designs, a Feedback ANC topology places its sense microphone (termed the Error Microphone) directly inside the acoustic cavity of the earcup, positioned immediately in front of the loudspeaker diaphragm and adjacent to the entrance of the user's ear canal.

Instead of predicting how external noise will penetrate the earpad, the feedback error microphone directly measures the combined residual acoustic pressure $e(n)$ inside the earcup—which represents the sum of the leaked ambient noise and the anti-noise generated by the headphone driver. The closed-loop controller continuously adjusts its output to drive the error signal $e(n)$ toward zero.

Because the error microphone sits downstream of the loudspeaker driver, the acoustic delay between the driver diaphragm and the error microphone sits directly inside the closed feedback loop. Control theory dictates that any phase lag inside a negative feedback loop reduces phase margin. When the cumulative phase lag of the driver, the internal air cavity, and the electronic filter reaches $180\text{ degrees}$ (converting negative feedback into positive feedback), the loop gain must be rolled off well below $0\text{ dB}$ ($1.0$) to satisfy the Nyquist Stability Criterion and prevent violent high-frequency howling oscillation.

Furthermore, because the error microphone picks up both unwanted ambient noise and the desired music playback signal, the feedback controller must subtract the known digital music signal (filtered through an estimate of the driver-to-microphone secondary path $\hat{S}(z)$) from the error microphone signal before passing it into the anti-noise feedback filter. Without this music-subtraction loop, the feedback ANC system would treat the bass frequencies of the user's music as "noise" and aggressively cancel out the low end of the audio track.

Flagship active noise-cancelling headphones and true wireless in-ear monitors universally implement a Hybrid ANC architecture. Hybrid ANC deploys at least two dedicated microphones per earcup: an outward-facing feedforward reference microphone and an inward-facing feedback error microphone operating simultaneously through a dual-core low-latency DSP engine.

In an advanced adaptive hybrid architecture, the internal error microphone accomplishes two tasks simultaneously: it runs a tight, low-frequency closed-loop feedback controller to crush sub-bass rumble ($20\text{ Hz – }300\text{ Hz}$), AND it feeds its residual error signal $e(n)$ into an adaptive Filtered-X Least Mean Squares (FxLMS) algorithm that dynamically updates the FIR filter coefficients of the outer feedforward filter $W(z)$ in real time. If the user puts on glasses or turns their head, the FxLMS engine detects the changed leakage pattern via the error microphone and recalibrates the feedforward filter within milliseconds.

Why can't a standard Least Mean Squares (LMS) adaptive filter be used directly in active noise cancellation? In traditional digital signal cancellation (such as telephone line echo cancellation), the filter output is subtracted electrically from the error node with zero delay. In acoustic ANC, however, the digital filter's output $y(n)$ must pass through a Digital-to-Analog Converter, a power amplifier, the non-linear electromagnetic voice coil of the headphone driver, the acoustic compliance of the air cavity, and the error microphone's transducer before it sums acoustically with the noise wave.

This electro-acoustic chain is known as the Secondary Path $S(z)$. It introduces substantial group delay and phase rotation. If a standard LMS algorithm updates its weights $w(n)$ using the raw input reference $x(n)$ correlated against the delayed acoustic error $e(n)$, the phase mismatch causes the adaptive algorithm to diverge and go unstable. The Filtered-X LMS (FxLMS) algorithm solves this by filtering the reference signal $x(n)$ through an internal digital model of the secondary path $\hat{S}(z)$ before computing the gradient descent weight update:

1. Anti-Noise Filter Output (N-tap FIR Filter):
   y(n) = sum_{k=0}^{N-1} w_k(n) * x(n - k) = w^T(n) * x(n)

2. Filtered Reference Signal (Passing x(n) through Secondary Path Estimate S_hat):
   x'(n) = sum_{m=0}^{M-1} s_hat_m * x(n - m)

3. Stochastic Gradient Descent Weight Update Rule:
   w(n + 1) = w(n) + mu * e(n) * x'(n)

Where:
- w(n)   = Adaptive FIR filter tap weight vector at sample n
- mu     = Adaptation step-size parameter (controls convergence speed vs stability)
- e(n)   = Residual acoustic error measured by internal error microphone
- s_hat  = Impulse response of the estimated electro-acoustic secondary path

Modern flagship headphones perform continuous online secondary path modeling: by monitoring how the user's normal music playback (or an imperceptible sub-audible pilot tone) registers at the internal error microphone, the DSP tracks real-time changes in acoustic impedance inside the ear canal—automatically adjusting both the ANC filter weights and the playback equalization curve to match the exact acoustic volume of the listener's ears.

High-resolution audio playback DACs typically employ multi-bit delta-sigma modulators paired with sharp linear-phase finite impulse response (FIR) decimation/interpolation filters. While a 128-tap linear-phase FIR filter provides pristine passband flatness and zero phase distortion for music, it introduces a group delay of $30\text{ to }60\text{ samples}$—translating to over $1,000\text{ microseconds}$ ($1\text{ ms}$) of latency at a $48\text{ kHz}$ sampling rate. As established by the causality constraint, $1\text{ ms}$ of latency is more than ten times too slow for acoustic ANC.

Consequently, dedicated ANC coprocessor chips bypass conventional audio converters entirely. Instead, they utilize ultra-high-oversampling Successive Approximation Register (SAR) ADCs or low-order Delta-Sigma modulators clocked at $768\text{ kHz}$ to $3.072\text{ MHz}$, paired with minimum-phase Infinite Impulse Response (IIR) biquad filter cascades or asynchronous decimationless processing cores. By operating IIR biquad filters directly at $768\text{ kHz}$, total round-trip ADC-to-DSP-to-DAC electronic latency is compressed down to 3 to 8 microseconds, leaving over $90\%$ of the physical time-of-flight budget available for phase-accurate anti-noise synthesis.

Sealing the ear canal or circum-aural cavity introduces two well-known psychoacoustic and physiological phenomena that must be actively managed by the headphone's acoustic and DSP design:

• The Osteophonic Occlusion Effect: When the entrance to the ear canal is plugged by a closed-back headphone or silicone IEM tip, bone-conducted vibrations generated by the listener's own vocal cords, chewing, or footsteps vibrate the cartilaginous walls of the ear canal. Normally, these low-frequency vibrations escape out the open ear canal. When sealed, the trapped cavity acts as a pressure chamber, boosting internal sound pressure levels between $80\text{ Hz}$ and $400\text{ Hz}$ by up to $20\text{ to }30\text{ dB}$—making the user's voice sound hollow and booming inside their head. Feedback ANC loops actively suppress the occlusion effect because the internal error microphone detects this bone-conducted pressure buildup inside the cavity and generates anti-phase waves to cancel it out.
• "Eardrum Suck" (Psychoacoustic Barometric Illusion): Many users report feeling an uncomfortable sensation of physical air pressure on their eardrums when switching ANC on, even in a quiet room. Physical manometer measurements confirm that ANC headphones produce zero static barometric (DC atmospheric) pressure change. Instead, this sensation is a neural perceptual illusion: the human brain relies on the continuous, ubiquitous background presence of sub-bass ambient room noise ($15\text{ Hz – }60\text{ Hz}$) as a baseline reference. When an aggressive feedback ANC loop suddenly strips away $30\text{ dB}$ of low-frequency ambient energy while leaving high-frequency earpad pressure intact, the brain misinterprets the sudden absence of low-frequency tympanic membrane motion as a static pressure differential across the eardrum (identical to what occurs during aircraft cabin ascent when hydrostatic pressure stiffens the eardrum).

To allow users to converse and maintain situational awareness without removing their headphones, modern ANC systems implement Transparency Mode (Hear-Through). Rather than turning the external microphones off, Transparency Mode repurposes the exterior feedforward microphone array to capture external sounds and mix them into the headphone driver. However, simply amplifying the external mic sounds unnatural because the physical headphone shell blocks high frequencies while letting low frequencies leak through. A reference-grade Transparency DSP filter applies an inverse passive-insertion-loss equalization curve—injecting high-frequency air and mid-band vocal presence in exact phase alignment with the low-frequency sound leaking naturally through the earcup, while using dual-microphone beamforming to preserve natural front-to-back directional localization cues.