Why architecture matters here
The architecture matters because howling is a stability problem, and stability problems do not degrade gracefully — they cross a threshold and blow up. Below the critical loop gain the system is fine; a hair above it and a tone grows without bound in a fraction of a second. This binary, cliff-edged behavior means the goal is not to 'reduce howling' as if it were noise but to increase the stability margin — to raise the loop gain at which the system would go unstable, so the operator can run the useful gain they need with room to spare. The industry metric for this is maximum stable gain (MSG), and the value a suppressor adds is added stable gain (ASG): how many extra decibels of stable amplification the technique buys. Everything is measured against that.
The core reason multiple mechanisms exist is that the feedback loop can be attacked at different points, and each mechanism has a different cost/benefit. Notch filtering is surgical and cheap but reactive — it waits until a frequency starts ringing, then places a narrow notch there; it can chase howls around the spectrum and it removes real signal at the notched frequency. Frequency shifting is elegant and broadband — shifting the whole signal a few hertz means the returning energy no longer coincides with the outgoing energy at the same frequency, so the loop cannot build coherently — but too much shift audibly detunes the audio, so it is limited to small shifts that buy only modest ASG. Adaptive feedback cancellation (AFC) is the heavyweight: it continuously estimates the acoustic path from speaker to mic and subtracts a model of the fed-back signal, which can buy far more ASG but is computationally heavy and prone to a specific, nasty failure — cancelling correlated program audio.
The second architectural pressure is distinguishing feedback from wanted audio, and it is genuinely hard because a howl and a sustained musical note look similar in the short term: both are narrowband, both persist, both are loud. A detector that notches anything narrowband and persistent will gut a held organ chord or a flute; a detector that is too conservative lets the howl establish itself before acting, by which point it is already loud. The design has to exploit the features that separate them — a howl's frequency is locked to the room's acoustics and grows monotonically, and it tends to recur at the same frequencies, whereas music moves and has harmonic structure — and it has to decide fast, because a howl that is allowed to grow for even a second is an unpleasant experience for everyone in the room.
The third pressure is latency. All of this runs live, in the path between microphone and loudspeaker, and any processing delay is delay the audience hears and, worse, delay that changes the loop's phase behavior. The suppressor's own latency is part of the system it is trying to stabilize, so every millisecond of algorithmic delay is scrutinized, and the detection and adaptation have to be fast enough to catch a howl in its first few cycles rather than after it has fully bloomed.
The architecture: every piece explained
Top row: detection and the reactive suppressors. The mic input is analyzed continuously in the frequency domain (a running FFT). The feedback detector looks for the telltale signature of an incipient howl: a spectral peak that is narrow, rising, and — the discriminating feature — persistent and recurrent at a room-locked frequency, often flagged by peak-to-harmonic-power ratios or by the fact that the peak keeps returning at the same bin. When it fires, a notch filter is placed precisely at that frequency: a narrow, deep attenuation that removes the loop gain exactly where the instability is, ideally narrow enough that little wanted audio is lost. In parallel, a small frequency shift (typically a few hertz, imperceptible on speech) decorrelates the loop broadband, so returning energy never perfectly reinforces outgoing energy — a cheap, always-on margin booster that complements the reactive notches.
Middle row: the adaptive core and gain. The adaptive filter (AFC) is the powerful mechanism: it maintains a model of the feedback path — the impulse response from loudspeaker to microphone through the room — and continuously updates it (via LMS/NLMS-style adaptation), subtracting its prediction of the fed-back signal from the mic input so the loop never sees it. Because the room path changes as people move and doors open, the model must adapt, but adapting while program audio is playing is exactly where it can go wrong (see failure modes). The gain manager is the safety net: when the howl-margin estimate says the system is near instability, it reduces gain — a blunt but reliable last resort. The output mix goes to the loudspeaker, and the room / air path is the physical feedback loop the whole system is fighting — the thing the AFC is trying to model and cancel.
Bottom rows: the intelligence that keeps it musical. The howl-margin estimate quantifies how close the system is to instability, driving both the gain manager and how aggressively the other stages act. The music-vs-howl classifier is what prevents the suppressor from destroying legitimate audio: it uses temporal and harmonic features to tell a growing room-locked tone from a sustained musical note, so a held flute note is passed while a howl at the same frequency is notched. The ops strip is the scorecard: added stable gain (how much extra headroom the suppressor bought), the latency budget (kept tight for live use and loop stability), and artifact monitoring — because over-notching, detuning, and AFC misadaptation all leave audible fingerprints that must be watched.
End-to-end flow
Walk a conference-room speakerphone through a howl event. A remote participant's voice plays out the room loudspeaker. The local gain is set high so people at the far end of the table are heard clearly — comfortably useful, but not far from the room's stability limit. Someone picks up the handheld mic and walks toward the loudspeaker to point at a slide. As they close the distance, the acoustic path from speaker to mic strengthens; the loop gain at some room-resonant frequency creeps toward unity.
At the microphone, energy at that frequency begins to build. The feedback detector, running its FFT, sees a peak that is narrow and — critically — growing and recurring at the same bin over successive frames, a signature no speech or music produces. Before the peak is loud enough for a human to register as a screech, the detector fires. The music-vs-howl classifier confirms it is not a sustained note (no harmonic partners, room-locked frequency, monotonic growth), and a notch filter is placed at that exact frequency. The loop gain there drops below unity, the growth reverses, and the tone dies in a few tens of milliseconds — most people in the room never consciously heard it. Meanwhile the always-on frequency shift has been quietly denying the loop the perfect coherence it needs, buying a couple of decibels of margin the whole time.
Behind all this, the adaptive feedback canceller has been modeling the room path continuously. As the person walked toward the speaker, the path changed, and the AFC updated its estimate, subtracting its prediction of the fed-back audio from the mic so the useful loop gain stayed high without the feedback contaminating it. The combination is what matters: the AFC raised the baseline stable gain, the frequency shift added standing margin, and the notch handled the specific tone that broke through — three mechanisms, each covering the others' weaknesses.
Now the stress cases. Suppose a remote participant plays a music clip with a long sustained note that happens to sit at a room-resonant frequency. A naive detector would notch it, gutting the music. Here the classifier recognizes the harmonic structure and steady (not growing) amplitude and holds off, passing the note. Suppose instead the AFC, adapting during that correlated music, starts mistaking the sustained note for feedback and cancelling it — 'entrainment,' the classic AFC failure. The system detects the misadaptation (the canceller's output correlates with program audio in a way real feedback would not), freezes or slows adaptation during highly correlated passages, and relies on notching and frequency shift for margin until the audio decorrelates. And if all else is losing — the room is simply too live and the gain too high — the gain manager pulls the master gain down a few decibels: less loud, but stable, which is always the right trade over a screaming room.