Quiet-Window Cumulants for Underdetermined Mixtures of Speech and Flat Signals

When more signals are mixed than there are sensors, for example three sources picked up by two microphones, the standard tools for pulling them apart share a hidden bias. Methods that rely on sparsity assume every source is "peaky": mostly near zero, with occasional large bursts, as speech is. Methods that rely on higher-order statistics, meaning averages of fourth powers that detect departures from a bell curve, can in principle handle "flat" sources too. Flat sources include digital communication signals, hum, and uniform-like interference. The catch is that these statistics are noisy, and the noise comes mostly from the peaky sources' rare, huge values. This note argues that in underdetermined mixtures the flat sources are the ones lost first when data are limited. It also argues for a cheap fix: fit the flat sources only on the stretches of time when the peaky sources are quiet. During those stretches the problem stops being underdetermined, and the sample averages are far less noisy.

The practical stakes are concrete. Hum in brain recordings, a modem tone under speech, and a jammer in a radio array are all flat sources that sparse separators discard or misclassify. The fourth-order methods that could recover them are known to need tens of thousands of samples before they stabilize, as reported for cumulant-based structure learning in Identification of Causal Structure with Latent Variables Based on Higher Order Cumulants. The proposal turns a property of speech that sparse methods already exploit, its frequent silences, into a way to cut the sample cost of estimating the flat sources.

Core Claim

In the model $x = As + n$, $x$ holds the $M$ sensor readings, $A$ is an unknown $M \times K$ mixing matrix with $K > M$, $s$ holds $K$ independent sources, and $n$ is Gaussian sensor noise. The fourth cumulant (excess kurtosis) of the projection $w^\top x$ equals $\sum_i \kappa_i (w^\top a_i)^4$. Here $a_i$ is the $i$-th column of $A$ and $\kappa_i$ is source $i$'s fourth cumulant: positive for peaky sources, negative for flat ones. The empirical estimate of this quantity has variance of order $\mathbb{E}[(w^\top x)^8]/N$ over $N$ samples, because it is a sample mean of fourth powers. Resolving column $i$ therefore needs roughly

$$N_i \gtrsim \frac{\mathbb{E}[(w^\top x)^8]}{\kappa_i^2}.$$

The numerator is dominated by the heaviest-tailed source. Flat sources have small $|\kappa_i|$: a uniform source has $\kappa = -1.2$ and a binary one $-2$. Speech has large positive kurtosis and an enormous eighth moment. The prediction is an asymmetric failure: global cumulant methods recover the speech columns long before the flat ones.

Now restrict to the fraction $\rho$ of samples in which the peaky sources are inactive. On those samples the active sources are only the $K_{\text{flat}} \le M$ flat sources plus noise. The subproblem is determined, and the eighth moment in the numerator collapses to that of light-tailed signals. The predicted gain is the ratio of the full-data eighth moment to $\rho$ times the quiet-window eighth moment. For speech with a realistic silence fraction this should be large. It should shrink, and eventually reverse, as $\rho \to 0$. The central claim to test is this crossover: a gated estimator reaches a given column error for the flat sources with several-fold fewer total samples than global overcomplete cumulant decomposition. The gain should track the predicted eighth-moment ratio, and it should disappear when the peaky sources never fall silent.

Research Program

The estimator has three stages.

The first stage estimates the columns of $A$ belonging to peaky sources by the standard single-source-point idea. In a short-time Fourier representation, bins dominated by one sparse source line up along that source's column. These are clustered, as in the DUET approach used in Sparse component analysis for speech recognition in multi-speaker environment. This step is sample-cheap and needs no higher-order statistics.

The second stage builds the gate. For each frame, the method scores peaky-source activity from the energy of the time-frequency points assigned to those columns, smoothed over a few hundred milliseconds. Frames below a threshold are kept.

The third stage runs a noise-invariant, sign-agnostic fourth-order iteration on the kept frames. The iteration is the pseudo-Euclidean gradient iteration (PEGI) of A Pseudo-Euclidean Iteration for Optimal Recovery in Noisy ICA. That method handles sources of opposite kurtosis signs without switching rules and ignores Gaussian noise, but it is defined only for determined mixtures. Gating is what makes its determined assumption hold, and that is the sense in which this proposal extends it. Waveforms are then recovered with the full estimated $A$. The first recovery rule is the demixing matrix $\widetilde{A}^\top \mathrm{cov}(x)^\dagger$, which the PEGI paper proves maximizes signal-to-interference-plus-noise ratio even when $K > M$ but never exploits algorithmically. The second is a maximum-a-posteriori (MAP) recovery using a Laplacian prior for peaky sources and a flat prior for the others.

The main risk is selection bias. Gating on observed energy also conditions on the flat sources' values, which distorts their cumulants. The design choice against this is to gate on slow, frame-level activity of the peaky sources rather than on per-sample magnitude. Flat sources are typically stationary within a frame, so frame-level gating should weakly condition them. This is an empirical hypothesis, and the program measures the bias directly. It compares cumulants of the flat sources before and after gating, and it compares the gate against an oracle gate built from the true speech activity.

The theory component is modest and specific. It bounds the flat-column error of the gated estimator in terms of $\rho$, the leakage of peaky energy into kept frames, and the quiet-window eighth moment. It then contrasts that bound with the variance expression above for the global estimator. The target is a crossover condition, not a general overcomplete sample-complexity theorem.

Baselines

The naive reference point is linear-programming or $\ell_1$ sparse recovery after clustering. This treats every source as peaky. In the $2\times 3$ speech-plus-flat setting of Blind separation of speech and sub-Gaussian signals in underdetermined case, this baseline gave separation error $e_i = 1.92$ on the flat source and misclassified its kurtosis sign. The same data and clustering stage will be used as in the proposed method, so the only difference is the missing gated stage.

The first strong competitor is that paper's own method. It invents $K - M$ "hidden observations" to complete the mixture and runs Extended Infomax, whose nonlinearity switches with the estimated kurtosis sign. It reports $e_i = 0.15$ on the flat source in that setting. That figure is the number to beat on the same synthetic construction. It will be re-run across a sweep of sample sizes and noise levels, since the paper gives neither. Beating it means lower flat-source error at matched $N$, and graceful behavior as $K - M$ grows beyond one and as sensor noise increases. The paper does not address either condition.

The second strong competitor is global overcomplete cumulant decomposition. One version is Fourier PCA from Fourier PCA and robust tensor decomposition, which provably learns fat $A$ under Gaussian noise; no reported result found on Lacuna for its empirical accuracy. The other is canonical polyadic tensor initialization followed by NMF-EM from Underdetermined Blind Source Separation Combining Tensor Decomposition and Nonnegative Matrix Factorization. That method reports SDR "around 25-30 dB" on SiSEC'08 linear music mixtures versus 10–15 dB for older baselines, but on mixtures with no flat sources. These methods are the direct test of the core claim, because they use all the data and all the fourth-order information. All methods get the same samples, the same tensor order, and the same hyperparameter search budget, with the search done on held-out mixtures. Beating them means matching their peaky-column accuracy while reaching their flat-column accuracy with at least a few-fold fewer samples.

Two ablations isolate the key ingredient. The first replaces the learned gate with a random subset of the same size, which removes the tail-variance reduction while keeping the sample count. The second replaces it with an oracle gate, which removes gate error. If random subsampling matches the learned gate, the claim is false.

Why not just…?

Why not just run a global overcomplete tensor decomposition on all the data? This is the real threat. Using every sample could outweigh the variance reduction if silences are rare. The eighth-moment argument predicts where each approach wins. The sweep over silence fraction $\rho$ will settle it: the gated method should win at natural speech silence rates and lose as $\rho \to 0$. If it never wins, the idea is wrong.

Why not just cluster time-frequency points, DUET-style, for all sources? Flat sources occupy the time-frequency plane continuously and so violate the one-source-per-bin assumption. The $\ell_1$ baseline above failed on exactly this ($e_i = 1.92$ in Blind separation of speech and sub-Gaussian signals in underdetermined case). The proposal keeps clustering for what it does well, the peaky columns, and replaces it only where it fails.

Why not just use second-order structure, as in sparse ICA? On the Identifiability of Sparse ICA without Assuming Non-Gaussianity identifies $A$ from covariance alone with a few hundred samples. It does so only for square, triangularizable mixing matrices, and covariance is biased by additive noise, as the Gaussian-noise-invariant line of Fast Algorithms for Gaussian Noise Invariant Independent Component Analysis shows. The proposed method will report results with sensor noise up to signal level, where covariance-only estimates should degrade and gated cumulants should not.

Why not just gate on raw mixture energy per sample? It is simpler, but it conditions directly on the flat sources and should bias their cumulants toward zero. This variant is run as an explicit ablation. If it performs as well as frame-level activity gating, the simpler version will be adopted.

Proposed Experiments

The first experiment is synthetic and reproduces the $2\times 3$ construction of the Kim and Yoo paper: two speech sources and one uniform or binary source. It then extends to $3\times 5$ and $4\times 7$ with one to three flat sources. The sweep covers $N$ from $10^3$ to $10^5$, Gaussian noise from 0 to 100% of signal variance, and silence fraction $\rho$ controlled by splicing. The reported quantities are per-column angular error of $A$, separately for peaky and flat columns, the flat-source error $e_i$ in the Kim and Yoo definition, and output signal-to-interference-plus-noise ratio. The decisive check has three parts. Global methods should show the predicted asymmetry, with flat columns failing at larger $N$ than peaky ones. The gated method's sample advantage should track the predicted eighth-moment ratio. That advantage should vanish at $\rho \approx 0$.

The second experiment uses SiSEC'08 development datasets A and B, the linear instantaneous music and speech mixtures, with an added synthetic flat interferer. SDR, SIR, and SAR are reported against the tensor-NMF pipeline. The goal is to show that adding a flat source hurts that pipeline's flat-source recovery while gating restores it without lowering the 25–30 dB regime on the original sources.

The third experiment is a real-data check on multichannel EEG with power-line interference. Extended ICA Removes Artifacts from Electroencephalographic Recordings shows that ordinary ICA misses this interference while sign-adaptive ICA recovers it. Channels are subsampled until the mixture is underdetermined. The full-channel sign-adaptive decomposition serves as reference, and the metric is recovery of the line-noise component from few channels and short recordings.

The proposal fails if random subsampling matches the learned gate, if the oracle gate gives no gain over global decomposition, or if the measured selection bias on flat-source cumulants exceeds the variance saved.