Dereverberation, Echo Cancellation, and Acoustic Feedback

Three things go wrong with sound in a room, and they arrive sounding alike. Reverberation is the room's extended response to a source. Echo is delayed playback or a reflection, and it may come with a reference signal — a copy of what was played, already sitting in memory. Acoustic feedback is a loop, and a loop can become unstable. Different structures leave you different things to measure. One generic suppression model is rarely enough.

Feedback alone has half a century of control theory behind it. The survey of it runs to 40 pages in Proceedings of the IEEE, published in 2011 by van Waterschoot and Moonen. Its abstract sets out the scope: “The acoustic feedback problem has intrigued researchers over the past five decades, and a multitude of solutions has been proposed. In this survey paper, we aim to provide an overview of the state-of-the-art in acoustic feedback control, to report results of a comparative evaluation with a selection of existing methods, and to cast a glance at the challenges for future research.” Fifty years of dedicated method development, and a comparative evaluation of the methods against each other. No field builds that for a special case of noise suppression.

Only one of the three hands you a reference, and echo cancellation is built on it. It assumes a usable reference, and a path that can be tracked. Someone moving the device, nonlinear loudspeakers, clock mismatch, double-talk and feedback each violate one half of that model. Which architecture you pick matters less than whether you named the problem correctly first. Suppression aimed at the wrong phenomenon is still suppression.

Two things do the naming: when the unwanted sound arrives, and whether it correlates with something you already hold. Playback echo is delayed, and it correlates with the digital reference that produced it. Reverberation is the room's own decay. It correlates with the source, but there is no clean copy of the source to subtract. Feedback returns the output to the input, so what it correlates with is the system's own past. That is what makes it a loop that can grow rather than a response that dies away.

That test has to come first, for a structural reason. The work runs in three stages: identify the phenomenon, map the available references, control residual and failure. They are ordered by dependency, not by convenience. Mapping the references consumes the answer the first stage produced. It can inherit the label, not question it. Only residual and failure control puts the label under strain, and it sits at the end. By then the suppressor has already been built around whatever name it was given.

A meeting device made exactly that mistake. It treated far-end playback — the far end's audio, coming back out of its own loudspeaker — as ordinary background noise. Everything needed to name it correctly was present. The playback correlated with a known digital reference, and it changed with the room acoustics. Filed as noise, none of that was used. The denoiser left residual echo. During double-talk it sometimes removed near-end speech instead, cutting off the person in the room to quieten a signal it had misnamed.

Averaged echo-removal figures would have looked acceptable here. That is how a team ends up suppressing what is left over before the canceller has settled, with nobody reading the one number that would have shown it. The figures later in this lesson are the field's own measurements of that gap — between a removal average and a listener.

What separates the three problems is what each one gives you to aim at.

Dereverberation reduces late room energy with no perfect dry reference to aim at. That is not an inconvenience of practice. It is written into the benchmark. The REVERB challenge, from Kinoshita and eleven colleagues in 2013, fixes the conditions and then withholds the room: “SimData simulates 6 different reverberation conditions: 3 rooms with different volumes (small, medium and large), 2 types of distances between a speaker and a microphone array (near=50 cm and far=200 cm). The reverberation times (T60) of the small, medium, and large-size rooms are about 0.25, 0.5, 0.7 s, respectively.” Stationary noise is added at 20 dB SNR. Alongside the simulations sit real MC-WSJ-AV recordings, made in a meeting room with a T60 of 0.7 s. The rules explicitly forbid entrants the room's own parameters — the reverberation time, the direct-to-reverberation ratio, the impulse responses themselves. The dry reference is absent by design. A method that quietly leans on it is disqualified, not merely optimistic.

Acoustic echo cancellation and feedback control work from other information, and each has its own literature to be measured against. The feedback survey does not only describe the methods; van Waterschoot and Moonen report “results of a comparative evaluation with a selection of existing methods”. That is what it means for a problem to be theorised in its own right. The meeting device is what happens when the three are not diagnosed in their own terms. It held the most usable information of the three, a digital copy of the offending signal. It spent it as though it had none.

Diagnosing an acoustic loop is the exercise, and the meeting device is the shape of it. Write it up so another team can argue with the point at which you started suppressing residual echo, and with your evidence that the linear estimate had settled by then. That ordering is the claim most worth attacking, and the one most write-ups leave undefended. Record the states you tested in the terms the benchmarks use, so the write-up can be compared with something: far-end only, near-end only, double-talk, movement, silence. For the reverberant slice, state a T60 rather than the word “live”. Then set down what your diagnosis assumes. Give one counterexample that would break it. Say what you would change in the suppression once you know.

A diagnosis ends in numbers, and the meeting device is the reason no single one of them will do. Its removal average looked fine while the fault was plainly audible. The challenge organisers measured that gap rather than asserting it. On single-talk echo scenarios, ERLE correlated with subjective ITU-T P.808 ratings at PCC 0.31 and SRCC 0.23, against PCC 0.67 and SRCC 0.57 for PESQ. They also state where the figure is defined at all: “ERLE is only appropriate when measured in a quiet room with no background noise and only for single talk scenarios (not double talk), where we can use the processed microphone signal as an estimate for e(n).” A quiet room, single talk. That is the slice the meeting device passed, and the only slice on which its headline number was even meaningful.

Four kinds of evidence sit behind a release claim, and they answer four different questions. Report all four on the difficult slices as well as the clean ones. The slices where suppression ran ahead of the canceller are exactly the ones an average covers up.