Asymmetric COLA Window Optimization

• The paper introduces OLA-DPSS windows that optimize spectral energy concentration under quadratic COLA constraints, achieving improvements of up to 2–3 dB in side-lobe reduction.
• Asymmetric COLA windows enable low-latency STFT processing and effective dynamics control by decoupling analysis and synthesis window functions while maintaining perfect reconstruction.
• Minimum-phase approaches yield reduced intrinsic latency by front-loading energy concentration, with redesigned dual windows restoring exact COLA properties.

Asymmetric constant overlap-add (COLA) window optimization is the design of finite-support window functions whose shifted copies satisfy an overlap-add perfect-reconstruction or partition-of-unity constraint while the window shape is allowed to be non-symmetric in time. In audio and speech processing, the problem arises in at least three closely related settings: short-time Fourier transform (STFT) analysis/synthesis, low-latency analysis–synthesis with different analysis and synthesis windows, and overlap-add interpolation of framewise gain trajectories in dynamics processing. Across these settings, the central requirement is that overlap-add behavior remain exactly controlled, whereas the optimization target may be spectral energy concentration, smoothness, attack/hold/release dynamics, or intrinsic latency (Bäckström, 2019, Wang et al., 2021, Luo et al., 9 Jul 2025, Su et al., 2016).

1. Reconstruction constraints and the meaning of asymmetry

In the overlap-add framework used for speech and audio processing, windowed frames are processed independently and then recombined by shifting and summing them. For the standard 50% overlap case with hop size $R=L/2$, a real finite-length window $w_k$ used for both analysis and synthesis satisfies perfect reconstruction and uniform error variance when it obeys the Princen–Bradley condition

$w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$

This is a squared COLA condition. In the more general square-root form, perfect reconstruction is expressed as

$\sum_h w^2[n-hR] = C,$

while for distinct analysis and synthesis windows $w_a$ and $w_s$ the relevant condition is

$\sum_h w_a[n-hR]\,w_s[n-hR] = C.$

These equations constrain overlap structure, not time-reversal symmetry; asymmetry is therefore compatible with COLA so long as the overlap sums remain constant (Bäckström, 2019).

This distinction is operationally important. Symmetric windows are common because they are analytically convenient and often have desirable spectral behavior, but the reconstruction algebra itself does not require $w[n]=w[L-1-n]$. In the asymmetric analysis–synthesis construction used for low-latency speech separation, the pointwise product of analysis and synthesis windows is forced to equal a Hann prototype over the active synthesis support, and this product-level condition is sufficient for perfect reconstruction under hop $M$ even though the two windows are different and the analysis window is explicitly asymmetric (Wang et al., 2021).

2. DPSS-based COLA optimization and its asymmetric extension

A major line of work formulates overlap-add window design as an energy-concentration problem. The objective is the classical Slepian criterion: maximize the fraction of window energy that lies in a specified baseband. For a discrete window vector $\mathbf w\in\mathbb R^L$, this becomes the Rayleigh quotient

$w_k$0

where $w_k$1 is a symmetric Toeplitz matrix determined by a main-lobe parameter $w_k$2. Without overlap-add constraints, maximizing $w_k$3 yields the discrete prolate spheroidal sequence (DPSS). The 2019 formulation embeds the Princen–Bradley constraints directly into the optimization:

$w_k$4

This is a quadratically constrained quadratic program (QCQP), solved in the paper with MATLAB’s Optimization Toolbox (2018a) and an interior-point algorithm; convergence is reported in seconds for practical audio window lengths (Bäckström, 2019).

The resulting windows, denoted OLA-DPSS, are optimized within the manifold of overlap-add windows rather than designed first and adapted afterward. This contrasts with the Kaiser-Bessel derived (KBD) window, which is described as an approximation of the discrete prolate spherical sequence adapted for overlap-add. For $w_k$5, the reported first side-lobe reduction is about $w_k$6 relative to half-sine and KBD windows. The reported energy concentrations are $w_k$7 for half-sine, $w_k$8 for KBD, and $w_k$9 for OLA-DPSS. For low-overlap windows with $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$0 and $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$1, the differences appear mainly from the second side-lobe onward, where OLA-DPSS is about $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$2 better than half-sine and about $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$3 better than KBD; the corresponding energy concentrations are $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$4, $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$5, and $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$6 (Bäckström, 2019).

A logical extension of this QCQP to asymmetric COLA design is to retain the quadratic overlap-add constraints and omit any symmetry constraint. In the 50% overlap case, the Princen–Bradley equations already relate only pairs of samples separated by $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$7. For general hop size $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$8, the same idea can be expressed with generalized diagonal matrices $w_{k+L/2}^2 + w_k^2 = 1,\qquad k=1,\dots,L/2.$9 enforcing

$\sum_h w^2[n-hR] = C,$0

which is equivalent to a constant squared overlap sum over all indices participating in the overlap. The 2019 framework does not explicitly develop this asymmetric case, but its constraint structure does not mathematically forbid it (Bäckström, 2019).

3. Convex asymmetric COLA windows for attack–hold–release dynamics

A second, explicitly asymmetric formulation appears in multichannel mixer–limiter design, where COLA windows are not used for spectral analysis but for constructing continuous per-sample gain envelopes from framewise optimal gains. Here the envelope for channel $\sum_h w^2[n-hR] = C,$1 is

$\sum_h w^2[n-hR] = C,$2

and the window must satisfy bounded support, non-negativity, and the COLA property

$\sum_h w^2[n-hR] = C,$3

These conditions ensure that the time-domain envelope is a convex combination of feasible framewise gain solutions, so samplewise limiter constraints remain satisfied after overlap-add (Luo et al., 9 Jul 2025).

In this setting, asymmetry is functionally motivated by limiter dynamics. The window is divided into attack, hold, and release intervals determined by onset times $\sum_h w^2[n-hR] = C,$4 and $\sum_h w^2[n-hR] = C,$5. Monotone rise, constant plateau, and monotone decay are imposed through first-order finite-difference constraints, while smoothness is maximized by minimizing the squared second-order finite differences. The optimization problem is

$\sum_h w^2[n-hR] = C,$6

subject to

$\sum_h w^2[n-hR] = C,$7

$\sum_h w^2[n-hR] = C,$8

Because the quadratic kernel $\sum_h w^2[n-hR] = C,$9 is positive semidefinite and all constraints are linear, the problem is a convex quadratic program (Luo et al., 9 Jul 2025).

This formulation treats symmetry as a special case rather than a default. The paper states that symmetric onset times $w_a$0 yield symmetric windows. It further reports that late attack-onsets $w_a$1 present identical asymmetric windows, early attack-onsets $w_a$2 exhibit piece-wise flat regions, and the remaining cases have a single flat section in the hold-interval. The reported solutions are scale-invariant with respect to the ratio $w_a$3. The objective is not spectral concentration but smoothness under dynamics and COLA constraints, which makes the formulation particularly suitable for time-directional control problems such as fast attack and slower release in limiting (Luo et al., 9 Jul 2025).

4. Asymmetric analysis–synthesis window pairs for low-latency STFT processing

In low-latency DNN-based speech separation, asymmetric COLA design appears as an analysis–synthesis window pair rather than a single self-dual window. The central motivation is that low-latency applications restrict algorithmic latency to roughly $w_a$4–$w_a$5, while short windows such as $w_a$6 yield poor frequency resolution. The proposed solution is to decouple analysis and synthesis: use a longer analysis window for spectral estimation and target construction, but retain a short synthesis window so that inference latency remains low (Wang et al., 2021).

The construction is built from a Hann prototype of length $w_a$7,

$w_a$8

The analysis window $w_a$9 has length $w_s$0, with initial zeros of length $w_s$1, a raised half-Hann middle segment, and a trailing segment

$w_s$2

The synthesis window $w_s$3 is nonzero only over the last $w_s$4 samples and is zero-padded to length $w_s$5. Perfect reconstruction is enforced by the pointwise product condition

$w_s$6

Accordingly, on the intermediate active region,

$w_s$7

while on the last segment $w_s$8 (Wang et al., 2021).

This asymmetric pair was evaluated with a speaker-independent deep clustering model on WSJ0 and a speaker-dependent mask inference model on Danish HINT. The paper reports “an improvement in separation performance of up to $w_s$9 in terms of source-to-distortion ratio (SDR) while maintaining an algorithmic latency of $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$0.” In the Danish HINT online mask inference setting, the low-latency symmetric baseline $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$1 gives $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$2 SDR, whereas the asymmetric $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$3 configuration gives $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$4. In the WSJ0 online deep clustering setting, symmetric $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$5 gives $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$6 and asymmetric $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$7 gives $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$8. A controlled comparison further indicates that the improvement is attributed to more informative, higher-resolution targets rather than to a higher-resolution input representation alone (Wang et al., 2021).

5. Minimum-phase asymmetrization and intrinsic latency

A distinct theory of asymmetric window design is based on intrinsic latency. For a discrete window $\sum_h w_a[n-hR]\,w_s[n-hR] = C.$9, the relevant quantity is the gap between the observation time and the energy-centroid-based estimation time:

$w[n]=w[L-1-n]$0

In discrete time this is the centroid of $w[n]=w[L-1-n]$1. Symmetric windows place this centroid at the center of the support, so their intrinsic latency is half the window length (Su et al., 2016).

The minimum-phase construction starts from a symmetric window and produces an asymmetric companion with the same magnitude spectrum:

$w[n]=w[L-1-n]$2

or, when zeros on the unit circle make the logarithm singular, an $w[n]=w[L-1-n]$3-perturbed version using $w[n]=w[L-1-n]$4. The theoretical result is that if the original window has some zeros outside the unit circle, then the minimum-phase companion has smaller intrinsic latency while preserving the same magnitude spectrum. The mechanism is front-loaded energy concentration: among all sequences with the same magnitude spectrum, the minimum-phase sequence concentrates the most energy near the beginning of the support (Su et al., 2016).

The empirical consequences are substantial. Any symmetric window has intrinsic latency $w[n]=w[L-1-n]$5. For the minimum-phase flat-top window, the paper reports $w[n]=w[L-1-n]$6 and $w[n]=w[L-1-n]$7. For an optimized two-term cosine family with $w[n]=w[L-1-n]$8, the reported latency is $w[n]=w[L-1-n]$9. For the ITU-T G.729 hybrid Hamming–cosine window, the reported latency is $M$0. The onset-detection study gives a concrete performance trade-off: a symmetric flat-top window reaches an F-score of about $M$1 at intrinsic latency $M$2, while the minimum-phase flat-top reaches about $M$3 at $M$4 (Su et al., 2016).

The paper does not explicitly develop COLA-preserving dual-window design for these minimum-phase windows. It states, however, that the magnitude spectrum is preserved and the support length is unchanged. This suggests a route to asymmetric COLA optimization: begin with a symmetric COLA-capable prototype, apply the minimum-phase transform to obtain an asymmetric analysis window with reduced intrinsic latency, and then recompute an overview or dual window so that overlap-add perfect reconstruction is restored. The dual-window step is a plausible implication rather than an explicit construction in the paper (Su et al., 2016).

6. Scope, misconceptions, and unresolved questions

A recurrent misconception is that COLA implies symmetry. The formal conditions do not support that view. Princen–Bradley constraints, square-root COLA sums, and dual-window overlap products all regulate overlap structure rather than time reversal. Symmetry is often a property of familiar window families, not a requirement of reconstruction theory (Bäckström, 2019, Wang et al., 2021).

A second misconception is that all asymmetric constructions solve the same optimization problem. They do not. OLA-DPSS windows optimize spectral energy concentration under quadratic overlap constraints; mixer–limiter windows optimize smoothness under COLA and attack/hold/release inequalities; the asymmetric speech-separation pair is analytically constructed from a Hann prototype to decouple analysis resolution from synthesis latency; minimum-phase windows optimize intrinsic latency while preserving spectral magnitude. These are different objective functions under different operational constraints, even when all are described as asymmetric COLA design (Bäckström, 2019, Luo et al., 9 Jul 2025, Wang et al., 2021, Su et al., 2016).

Several limitations are also explicit. The OLA-DPSS work assumes real-valued finite-length windows, focuses on square-root COLA at 50% overlap and low-overlap variants, and yields numerical rather than closed-form solutions. The low-latency speech-separation work chooses the $M$5 analysis length empirically through oracle SDR at fixed $M$6 synthesis length, rather than by solving a general window-shape optimization. The mixer–limiter work presents no closed-form solution for the optimal asymmetric windows, no explicit spectral analysis, and no listening tests comparing asymmetric and symmetric windows. The minimum-phase latency work does not itself preserve exact COLA when the original synthesis window is left unchanged; exact overlap-add would require a redesigned dual window (Bäckström, 2019, Wang et al., 2021, Luo et al., 9 Jul 2025, Su et al., 2016).

Taken together, the literature defines asymmetric constant overlap-add window optimization as a family of constrained design problems in which asymmetry is introduced only insofar as it serves a specific systems objective—reduced side-lobes, low algorithmic latency, limiter-style dynamics, or reduced intrinsic latency—while overlap-add algebra remains exact or explicitly controlled. The field has therefore moved from treating symmetry as a default design assumption to treating it as one admissible solution within a broader constrained optimization space.