Removing room echo from speech can be posed as a separation problem. The recorded signal is split into the direct speech and the late echo, and the algorithm pushes the recovered speech to look as "non-Gaussian" as possible: peaky, with occasional large values and many near-zero ones, the way clean speech looks. On the REVERB Challenge, the one benchmark in this area with both simulated rooms and real recorded rooms, this idea helps on real one- and two-microphone recordings. It then loses to the older Gaussian-assuming method, Weighted Prediction Error (WPE), on real eight-microphone recordings (A Semi-Blind Source Separation Approach for Speech Dereverberation). That paper offers no explanation. Its peakedness setting was tuned and checked mainly on simulated mixtures, so this inversion is a clean instance of a simulation-derived assumption failing on real hardware.
This note proposes to explain the inversion and fix it without labels. The hypothesis has two parts. First, a strongly peaky prior gives very large weight to quiet time-frequency cells. In real rooms those cells are dominated by sensor and ambient noise rather than speech. Second, a longer eight-channel prediction filter has many more coefficients that this misweighting can corrupt. The proposed method keeps the non-Gaussian formulation but estimates, separately for each frequency band of each recording, both the prior's shape and a noise floor. It does this from the recording's own statistics, with no word-error feedback. Shapes on the sub-Gaussian side are allowed, so bands dominated by steady hum are not forced into a speech-like prior. The method succeeds if it matches the better of WPE and fixed-shape non-Gaussian dereverberation at every channel count on real data, with no tuning on real data.
Core Claim
The two methods compared in the Aux-ICA paper are points on one family. That paper proves WPE is equivalent to auxiliary-function independent component analysis (Aux-ICA, an independence-maximizing separation method whose updates need no step size) when speech is modeled as Gaussian with time-varying power. In both methods the dereverberation filter is fit by reweighted least squares. Each time-frequency frame is weighted by a power of its current estimated speech magnitude.
We propose the weight
$$w_{t,f} = \left(|d_{t,f}|^2 + \sigma_f^2\right)^{(\beta_f - 2)/2},$$
where $d_{t,f}$ is the current direct-speech estimate in frame $t$ and frequency $f$, $\beta_f$ is a band-specific shape parameter, and $\sigma_f^2$ is a band-specific noise floor. Setting $\beta_f = 0$ gives a WPE-like power normalization. Values of 0.2 to 0.4 reproduce the super-Gaussian Aux-ICA setting. Values above 2 describe flat, sub-Gaussian bands such as a constant-amplitude hum. Both $\beta_f$ and $\sigma_f^2$ are fit by maximizing the likelihood of a generalized Gaussian density plus Gaussian noise on the current output, alternating with the filter update.
The claim has two halves, and both are testable. The diagnostic half says the 8-channel inversion comes from unfloored small-$\beta$ weights acting on noise-dominated cells, not from speech being less super-Gaussian in real rooms. The constructive half says blind per-band estimation of $(\beta_f, \sigma_f^2)$ recovers the WER-optimal operating point at every channel count. Recovering it is not the same as recovering it only on average.
Relation to Prior Work
The Aux-ICA dereverberation paper supplies the method and the motivating anomaly. It fixes one global $\beta$ and has no noise term, and we change both. It reports a word error rate (WER) reduction on 1-channel RealData from 19.41% to 16.90% relative to WPE, and a WPE advantage on 8-channel RealData that the file gives without a number.
The idea of letting the source model adapt to whether a source is peaky or flat comes from instantaneous-mixture ICA:
- An alternative switching criterion for independent component analysis (ICA) toggles the nonlinearity by the sign of kurtosis.
- Extended ICA Removes Artifacts from Electroencephalographic Recordings uses that switching to remove sub-Gaussian 60 Hz mains noise that fixed Infomax leaves in.
- Nonlinearity and separation capability: further justification for the ICA algorithm with mixture of densities shows that only loose density matching is needed for consistent separation.
- ICA based on Split Generalized Gaussian fits shape parameters by maximum likelihood.
All four work on instantaneous mixtures or EEG. None of them has been tested in convolutive, reverberant rooms with real microphones. We carry their adaptivity into the frequency-domain reweighting used for dereverberation, and the Xu et al. result is what makes coarse per-band shape estimation plausible.
Correction: Independent Component Analysis for Brain fMRI Does Indeed Select for Maximal Independence motivates the diagnosis. It attributes separation failures to a mismatch between the assumed density and the true one, not to a lack of sparsity.
Two lines of work constrain the design. A Pseudo-Euclidean Iteration for Optimal Recovery in Noisy ICA handles mixed-kurtosis sources without switching, but only under Gaussian noise and only on synthetic data, which is why we model the noise floor explicitly rather than assume it cancels. Robust Independent Component Analysis Using Quadratic Negentropy shows that fourth-moment contrasts are fragile to outliers, which is why we fit shapes by likelihood rather than by kurtosis.
Finally, A Single-Channel Speech Dereverberation Method Using Sparse Prior Imposition in Reverberation Filter Estimation and A Unified Bayesian Source Modelling for Determined Blind Source Separation are examples of the evaluation practice we are reacting to. The first uses Image-Method shoe-box simulation and the second uses semi-synthetic mixtures with measured room responses. Both papers acknowledge their reality gap.
Research Program
The work has three stages, all on REVERB SimData and RealData at 1, 2 and 8 channels.
The first stage is diagnostic. We sweep fixed $\beta$ from 0 to 0.4, with and without a fixed floor, at each channel count on both SimData and RealData. On RealData we record WER, and on both we record the speech-to-reverberation modulation energy ratio (SRMR, a non-intrusive measure of how much echo remains). On SimData we also record the parallel distribution statistics. On the outputs we measure the fraction of total filter-update weight carried by the lowest-energy decile of cells, and how that fraction changes with channel count. The hypothesis predicts three things: real 8-channel data puts far more weight on quiet cells than simulated 8-channel data; the WER penalty of small $\beta$ grows with channel count on RealData but not on SimData; and adding a floor removes most of that penalty.
The second stage builds the blind estimator for $(\beta_f, \sigma_f^2)$. It includes a sub-Gaussian branch for bands with flat amplitude distributions. Estimation happens per utterance, interleaved with the filter iterations, at a cost of one one-dimensional likelihood search per band per iteration.
The third stage tests whether the blind choices track the oracle. For each channel count we find the WER-optimal fixed $\beta$ on RealData, then compare it with the distribution of blind estimates.
Baselines
The naive reference is the unprocessed microphone signal fed to the same speech recognizer. It anchors how much any dereverberation helps.
WPE is the strongest established competitor at 8 channels. The cited paper reports it overtaking Aux-ICA there, but the file gives no figure, so we will report our own reproduction.
Fixed-$\beta$ Aux-ICA with $\beta = 0.4$ is the strongest competitor at 1 and 2 channels. Its reported result is the reduction of 1-channel RealData WER from 19.41% to 16.90% (A Semi-Blind Source Separation Approach for Speech Dereverberation).
Fairness rules apply to every method. All methods share the same short-time Fourier transform settings, prediction delay, filter length per channel count, iteration count and recognizer. Fixed-parameter methods may tune only on SimData development data, mirroring how they would be deployed. Beating the competitors means two things together. At 1 and 2 channels, the method must match or improve on fixed $\beta = 0.4$. At 8 channels, it must match or improve on WPE. Both conditions are judged with paired significance over RealData utterances.
The ablations isolate the key ingredients. One fixes the floor at zero but estimates $\beta_f$. One fixes $\beta$ globally but estimates the floor. One uses a single global estimate instead of per-band estimates. One forbids $\beta_f > 2$, which isolates the sub-Gaussian branch.
Why not just…?
Why not just tune $\beta$ on real data per channel count? This is the real threat. If an oracle-tuned global $\beta$ matches our method, the constructive contribution reduces to a tuning recommendation. Oracle tuning needs transcribed recordings from the target room and array, which a deployed system does not have. We therefore report the oracle as an upper reference, not as a competitor. The decisive question is whether blind estimation gets within the paired-test noise of the oracle. If it does not, we report the diagnosis alone.
Why not just use WPE whenever there are many microphones? A channel-count switch would recover the better of the two reported results, but only if the crossover point does not depend on the room and noise. The per-room sweep in the first stage tests this directly. If the crossover is stable across REVERB's conditions, the switch is a sufficient practical answer and we will say so. The per-band method would still be justified only if it beats both endpoints somewhere.
Why not just regularize or shorten the 8-channel filter? If the inversion comes from estimation variance alone, diagonal loading or fewer taps would fix it with $\beta$ unchanged. We include a filter-length and loading sweep for fixed $\beta = 0.4$ in the first stage. If that closes the gap, the weighting hypothesis is wrong, and the proposal falls back to a negative result about where the simulation-trained prior fails.
Decisive Checks
The diagnosis is supported if three things hold together on RealData: the floored fixed-$\beta$ variant removes most of the 8-channel deficit; filter regularization does not; and SimData shows neither effect. The constructive claim is supported if blind per-band estimation is statistically indistinguishable from the per-channel-count oracle $\beta$ at all three channel counts.
A secondary check covers the sub-Gaussian branch. Bands with estimated $\beta_f > 2$ should coincide with spectral lines visible in non-speech segments of the RealData recordings, and forbidding that branch should measurably raise WER only in recordings that contain such lines.
References
- A Semi-Blind Source Separation Approach for Speech Dereverberation
- An alternative switching criterion for independent component analysis (ICA)
- Extended ICA Removes Artifacts from Electroencephalographic Recordings
- Nonlinearity and separation capability: further justification for the ICA algorithm with mixture of densities
- ICA based on Split Generalized Gaussian
- Correction: Independent Component Analysis for Brain fMRI Does Indeed Select for Maximal Independence
- A Pseudo-Euclidean Iteration for Optimal Recovery in Noisy ICA
- Robust Independent Component Analysis Using Quadratic Negentropy
- A Single-Channel Speech Dereverberation Method Using Sparse Prior Imposition in Reverberation Filter Estimation
- A Unified Bayesian Source Modelling for Determined Blind Source Separation