Blind Oblique Projection (BOP)

• Blind Oblique Projection (BOP) is a geometric method for estimating successive relative transfer functions in reverberant, noisy, multi-source acoustic environments.
• It constructs an oblique projection operator that preserves known source subspaces while nulling candidate directions to isolate newly activated sources.
• Variants like BOPO-W enhance robustness by incorporating orthogonal augmentation and noise-aware extensions to improve estimation in low-SNR conditions.

Blind Oblique Projection (BOP) is a geometric method for successive relative transfer function (RTF) estimation in reverberant, noisy, multi-source acoustic environments. In the formulation used in recent speech and beamforming work, BOP is designed for scenarios in which sources activate one after another: the RTF of the first source can be estimated by standard single-source methods, whereas the difficulty is to estimate the RTF of a newly activating source during segments in which previously active sources remain present. BOP addresses this by constructing an oblique projection operator that preserves the subspace spanned by already known active sources while nulling a candidate direction; the projected output power is then minimized with respect to that candidate, and the minimizer is taken as the RTF of the new source (Gode et al., 2023, Gode et al., 6 Aug 2025).

1. Problem class and operational setting

BOP arises in the estimation of RTF vectors for microphone-array processing, where accurate RTFs are used in beamforming for noise and interference suppression. The canonical setting is a successive-speaker scenario. In the two-speaker case, the observation sequence is divided into three segments: noise only, first speaker plus noise, and first speaker plus second speaker plus noise. The first speaker is estimated from the single-speaker segment; the second speaker must then be estimated from the dual-speaker segment, where both speakers are active simultaneously (Gode et al., 2023).

A multi-source generalization treats the $K$-th source as newly activating while sources $1,\dots,K-1$ are already active. In that case the microphone short-time Fourier transform (STFT) vector is modeled as the sum of the new source, the previously active sources, and noise. The task is to recover the RTF vector of the new source while blocking the contribution of the already-active mixture only to the extent required by the oblique geometry; equivalently, BOP seeks a projection that leaves the previously known source subspace distortionless and makes the new source disappear in the ideal model (Gode et al., 6 Aug 2025).

This places BOP in a narrow but important niche. It is not presented as a generic blind source separation method. Rather, it is a successive-identification technique that exploits partial prior knowledge: earlier source RTFs are assumed available, and a later source is identified through an optimization over oblique projectors.

2. Projection geometry and signal model

In the two-speaker formulation, the per-frequency microphone observation is modeled as

$$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$

where $\mathbf{h}$ is the RTF of the second speaker, $\mathbf{g}$ is the RTF of the first speaker, and the reference entry of both RTFs equals $1$ (Gode et al., 2023). The corresponding covariance in segment $s$ is

$$\mathbf{R}_{\mathbf{y},s} = \mathbf{h}\phi_{x,s}\mathbf{h}^{\mathrm H} + \mathbf{g}\phi_{u,s}\mathbf{g}^{\mathrm H} + \mathbf{R}_{\mathbf n}.$$

The core object in BOP is the oblique projection operator. In the two-speaker case it is defined by

$$\mathbf{P}^{\angle}_{\mathbf g \bm\theta} = \mathbf g \bigl( \mathbf g^{\mathrm H}\mathbf P_{\bm\theta}^{\perp}\mathbf g \bigr)^{-1} \mathbf g^{\mathrm H}\mathbf P_{\bm\theta}^{\perp},$$

with residual maker

$$\mathbf P_{\bm\theta}^{\perp} = \mathbf I_M - \frac{\bm\theta\bm\theta^{\mathrm H}}{\bm\theta^{\mathrm H}\bm\theta}.$$

This projector satisfies

$1,\dots,K-1$0

so it preserves the known first-speaker direction and blocks the candidate direction $1,\dots,K-1$1 (Gode et al., 2023).

The multi-source version replaces $1,\dots,K-1$2 by the matrix $1,\dots,K-1$3. Then the projector is described as projecting onto the span of $1,\dots,K-1$4 while nulling $1,\dots,K-1$5. If $1,\dots,K-1$6, the newly activated source is blocked ideally while the already active sources remain (Gode et al., 6 Aug 2025).

The geometric content is central. BOP does not use an orthogonal projection onto a noise or interference subspace. It uses a non-orthogonal decomposition in which one subspace is preserved and another direction is nulled. This is why the method remains tied to known source structure even when several sources are active simultaneously.

3. Optimization criterion and RTF recovery

BOP estimates the unknown RTF by minimizing projected output power. In the multi-source formulation, the cost function is

$1,\dots,K-1$7

Under the high-SNR or noiseless assumption, the same criterion is written with $1,\dots,K-1$8 in place of $1,\dots,K-1$9 (Gode et al., 6 Aug 2025). The RTF estimate is obtained from

$$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$0

In the two-speaker case, the objective is

$$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$1

The trace expansion shows that the first-speaker term is constant after projection. Minimizing projected covariance power therefore amounts to choosing $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$2 so that the remaining contribution is minimized; in the ideal model that occurs when $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$3 aligns with the second-speaker RTF $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$4 (Gode et al., 2023).

This is the standard intuition behind BOP: the correct candidate is the one that causes the newly active source to vanish after oblique projection while leaving the already active source subspace intact. In that sense, BOP is an RTF-identification method implemented through an optimal blocking condition rather than through direct eigendecomposition of a covariance matrix.

4. Modeling assumptions, degeneracies, and limitations

The standard BOP derivation assumes a linear microphone array, time-invariant acoustic conditions or spatial stationarity, per-frequency STFT processing, sufficiently large STFT frames so that each speaker is represented by a rank-1 RTF model, and uncorrelated source and noise components (Gode et al., 2023). These assumptions are structural: they define the regime in which oblique blocking corresponds to source-direction identification.

A major limitation of the conventional formulation is its high-SNR character. In the two-speaker analysis, the noise term in the projected covariance was neglected in prior work, and BOP was therefore described as assuming a sufficiently large SNR (Gode et al., 2023). Later work on multi-source successive estimation repeats this point more broadly: the conventional BOP formulation ignores the noise covariance, and performance degrades in low-SNR conditions (Gode et al., 6 Aug 2025).

A second limitation is algorithmic. The conventional method solves the minimization iteratively by gradient descent, which is computationally expensive and may converge to local minima (Gode et al., 6 Aug 2025). A further difficulty is a plateau phenomenon. If the candidate vector lies in the null space of $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$5, the oblique projection degenerates into a standard orthogonal projection onto $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$6, and the cost becomes independent of $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$7. To remove this degeneracy, the conventional approach augments the known-source matrix by additional vectors until rank $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$8 is reached (Gode et al., 6 Aug 2025).

The choice of these additional vectors is itself nontrivial. Earlier BOP variants used random additional vectors to avoid plateaus, but later work notes that, when the rank-1 acoustic model is not exact, random vectors can worsen estimation. The same work also remarks that oblique projection may increase magnitude when the angle between the range and nulling subspaces is small, making the method sensitive to modeling mismatch (Gode et al., 6 Aug 2025).

5. Closed-form variants, orthogonal augmentation, and noise-aware extensions

Recent work develops three extensions intended to make BOP practical in reverberant and noisy online settings (Gode et al., 6 Aug 2025). The first is a closed-form replacement for iterative optimization. After augmenting $$\mathbf{y}_t = \mathbf{x}_t + \mathbf{u}_t + \mathbf{n}_t = \mathbf{h}x_{r,t} + \mathbf{g}u_{r,t} + \mathbf{n}_t,$$9 to a rank-$\mathbf{h}$0 matrix $\mathbf{h}$1, the estimator becomes

$\mathbf{h}$2

where $\mathbf{h}$3 denotes the principal eigenvector. Because $\mathbf{h}$4 has rank $\mathbf{h}$5, the orthogonal complement has rank $\mathbf{h}$6, only one eigenvalue is nonzero, and the principal eigenvector is the stationary point of the cost gradient (Gode et al., 6 Aug 2025).

The second extension replaces random additional vectors by orthogonal additional vectors derived from the minor subspace of the covariance matrix: $\mathbf{h}$7 This variant is called BOPO. Its rationale is that the dominant eigenvector of $\mathbf{h}$8 is associated with the active source, while the minor subspace tends to be orthogonal to that dominant direction. The stated effect is improved robustness when the rank-1 model is imperfect and when SNR is low (Gode et al., 6 Aug 2025).

The third extension incorporates explicit noise handling. Noise subtraction uses

$\mathbf{h}$9

yielding BOP-S and BOPO-S. Noise whitening uses

$\mathbf{g}$0

followed by closed-form estimation in the whitened domain and de-whitening of the result, yielding BOP-W and BOPO-W (Gode et al., 6 Aug 2025).

Method	Additional vectors	Noise handling
BOP	random	none
BOP-S	random	subtraction
BOP-W	random	whitening
BOPO	orthogonal/minor-subspace	none
BOPO-S	orthogonal/minor-subspace	subtraction
BOPO-W	orthogonal/minor-subspace	whitening

The same work also introduces a spatial-coherence-based online source counting method, because both conventional BOP and its closed-form variants require frame-by-frame knowledge of when a new source becomes active. The method uses generalized magnitude-squared coherence (GMSC), a whitening step based on the covariance of already active sources plus noise, temporal smoothing, and thresholding to declare source activation (Gode et al., 6 Aug 2025).

6. Empirical behavior and comparison with alternative estimators

In a two-speaker reverberant laboratory experiment, BOP was compared with covariance whitening using the undesired covariance matrix (CWu) and with covariance blocking and whitening (CBW). The reported finding was that CWu achieves about $\mathbf{g}$1 dB signal-to-interferer-and-noise ratio (SINR) improvement, BOP outperforms CWu at low input SNRs, BOP performs worse at high input SNRs, and CBW outperforms both across all tested SNRs. An example given for CBW is about $\mathbf{g}$2 dB average SINR improvement at $\mathbf{g}$3 dB SNR (Gode et al., 2023).

The conceptual difference between BOP and CBW is explicit. BOP blocks the unknown second speaker while preserving the known first speaker. CBW instead blocks the known first speaker by a residual-maker matrix, whitens the residual noise, and then estimates the second speaker by singular value decomposition. CBW is presented as independent of the first-speaker power spectral density and as not assuming a large SNR (Gode et al., 2023). This suggests that later work increasingly interpreted BOP as a useful but noise-sensitive baseline rather than as the final solution for successive RTF estimation.

A broader three-speaker study with real-world reverberant noisy recordings evaluated BOP and its variants with measured room impulse responses and BRUDEX noise. In the hardest $\mathbf{g}$4 dB SNR triple-source condition, BOPO-W improves median weighted Hermitian angle by $\mathbf{g}$5 and SINR improvement by $\mathbf{g}$6 dB relative to conventional BOP. Across conditions, orthogonal additional vectors outperform random ones, whitening outperforms subtraction, and BOPO-W is reported as the best method (Gode et al., 6 Aug 2025).

When the same study replaced oracle source-activity information by online source counting, performance was only slightly worse: the maximal degradation was about $\mathbf{g}$7 in weighted Hermitian angle and $\mathbf{g}$8 dB in SINR improvement (Gode et al., 6 Aug 2025). This indicates that the practical bottleneck in successive BOP-style estimation is not only the oblique-projection step itself, but also the reliability of online activation detection.

7. Terminological ambiguity and relation to other oblique-projection literatures

The acronym “BoP” is ambiguous across research domains. In algebraic topology, “BoP” refers to Pengelley’s $\mathbf{g}$9-local spectrum used in the stable decomposition of $1$0, with no relation to microphone-array RTF estimation or source blocking (Wilson, 2018). This usage is unrelated to Blind Oblique Projection despite the superficial similarity of the acronym.

Several other arXiv literatures use oblique projection without referring to BOP. In observer design for nonautonomous semilinear parabolic-like equations, an explicit output-injection operator is built from oblique projections between sensor and auxiliary subspaces. That framework is described as conceptually very close to BOP because it uses finite measurements, complementary subspaces, and an oblique geometry, but it is not presented as the canonical BOP framework; its objective is semiglobal exponential observer stabilization in an infinite-dimensional PDE setting rather than algebraic RTF recovery (Rodrigues, 2020).

Likewise, Kaczmarz methods with oblique projection for large overdetermined linear systems use non-orthogonal projection directions constructed from system rows to accelerate convergence. These methods are explicitly framed as KO, RKO, GRKO, or MWRKO, not as BOP, and they belong to iterative linear-solver theory rather than successive source estimation (Li et al., 2021, Wang et al., 2021).

Accordingly, Blind Oblique Projection is best reserved for the acoustic signal-processing method in which a newly activating source is identified by minimizing the power of an obliquely projected covariance while preserving already known source directions. Other uses of “oblique projection” may share geometric motifs, but they solve different inverse problems and rely on different modeling assumptions.