The asker is probing a real pain point: when acoustic echo cancellation (AEC) breaks down — reverberant rooms, double-talk, abrupt acoustic changes — commercial stacks like WebRTC's AEC3, Zoom, and Meet fall back to residual echo suppression that effectively half-duplexes the call (volume ducking, near-end gating). Their proposal: instead of ducking the whole spectrum, partition it. Allocate disjoint frequency sub-bands to the near-end and far-end speakers in real time, so they literally can't acoustically collide in the loop.

Why it's interesting: AEC fundamentally relies on estimating a linear echo path. When that estimate diverges, the residual is unbounded. Frequency-domain duplexing sidesteps adaptive filtering by enforcing orthogonality in a different domain — borrowing an idea from FDMA radio. The clever bit is that speech is sparse in frequency: voiced segments concentrate energy in formants below ~3.5 kHz, with gaps. So you don't need to discard half the band — you allocate based on instantaneous spectral occupancy.

Why it's hard:

• Intelligibility cost. Speech intelligibility (STI, PESQ) collapses if you remove arbitrary bands. The Articulation Index weights 1–4 kHz heavily. Any partition must preserve enough of each talker's formant structure that the remote listener still parses words.
• Switching artifacts. Band reassignment during double-talk produces musical noise and "warbling" — the same artifact that plagues spectral-subtraction noise suppressors. You'd need overlap-add with smooth gain trajectories, not hard masks.
• It doesn't kill echo, it constrains it. The far-end signal still couples acoustically into the mic in its assigned band. You'd still need cancellation or suppression — partitioning just simplifies the problem because the near-end speech is (by construction) absent from those bins, so suppression has no double-talk to confuse it.

Prior art to check: The closest published work is perceptually-weighted residual echo suppression (Valin et al., RNNoise lineage) and frequency-domain double-talk detection in AEC3. Sub-band AEC itself is standard (partitioned-block frequency-domain adaptive filter, PBFDAF). What the asker is describing is closer to cognitive-radio-style dynamic spectrum access applied to acoustic full-duplex — I'm not aware of it as a deployed AEC fallback, though academic "spectral duplexing" papers exist for hearing aids.

Direction: Prototype with WebRTC's APM. Tap the far-end render signal, compute its short-time spectrum, and on a double-talk detector trigger, apply a complementary mask to the near-end capture instead of a flat suppression gain. Compare PESQ/POLQA against AEC3's default suppressor under reverberant conditions (image-source simulator, RT60 = 0.6s).

Gotcha: If the far-end is also doing this, you've created a coupled control loop across the network with RTT-bounded stability. Allocation must be unilateral or negotiated out-of-band.