Determined Multichannel Blind Source Separation with Clustered Source Model

In real-world applications, the observation of a source signal of interest frequently suffers from interference. Therefore, it is essential to use signal processing techniques to extract latent sources from observed mixtures by multiple microphones [1, 2]. Generally, there are two different paradigms for source separation. One is beamforming, which extracts the components from mixed signals using direction information with spatial filtering techniques [3, 4, 5]. They usually assume that the geometry of array and the incidence angle of each source are known. Another paradigm is to perform multichannel blind audio source separation (MBASS) based on independent component analysis (ICA) [6], which exploits the statistical independence of source signals. This study focuses on the latter paradigm.

Generally, MBASS is conducted in the short-time Fourier transform (STFT) domain to address convolutive mixing. However, a significant challenge arises due to inner permutations, which can greatly affect separation performance. To tackle this issue, independent vector analysis (IVA) [7] was then adopted, and the majorization-minimization (MM) principle [8] was introduced to derive fast update rules for IVA [9]. Despite its effectiveness in handling permutations, IVA-based methods often overlook spectral structure information. To incorporate spectral structure, nonnegative matrix factorization (NMF) [10] was extended to multichannel cases for MBASS [11, 12, 13]. However, Multichannel NMF methods are computationally demanding and sensitive to parameter initialization. In response to these challenges, the independent low-rank matrix analysis (ILRMA) method was devised [14]. ILRMA enforces an interpretable low-rank structure constraint on the spectrogram and employs a rank-one relaxation for the spatial model.

To further improve performance, efforts have been directed towards enhancing the source model by generalizing the distribution of source signals and achieving initialization-robust performance [15, 16, 17, 18, 19]. Additionally, research has explored the integration of prior constraints into parameters associated with the source model to enhance MBASS performance [20, 21]. However, these NMF-based methods are typically inadequate to capture inter-channel dependencies and higher-order structures inherent in multichannel data. While Kitamura [14] also discussed a CPD-based source model, its spectral distinctiveness remains somewhat deficient.

In this paper, we propose a clustered source model for determined blind source separation. Utilizing nonnegative block-term decomposition (NBTD), the source model parameters are expressed as a summation of several components, each being the outer product of a vector and a matrix. By applying orthogonality constraints to the latent vectors of this decomposition, they gain a clear interpretation, revealing distinct clusters of sources.

Let $M$ and $N$ be the number of microphones and sources, respectively. The observed signal in the STFT domain can be expressed as:

where the subscripts $i$ and $j$ denote, respectively, the frequency and frame indices, $I$ and $J$ denote, the total number of STFT bins and time frames, $\mathbf{x}_{ij}=\left[\begin{array}[]{ccc}x_{1ij}&\dots&x_{Mij}\end{array}\right]^{\intercal}\in\mathbb{C}^{M}$ and $\mathbf{s}_{ij}=\left[\begin{array}[]{ccc}s_{1ij}&\dots&s_{Nij}\end{array}\right]^{\intercal}\in\mathbb{C}^{N}$ are, respectively, the vectors consisting of the STFTs of the observation and source signals, respectively, $\mathbf{c}_{nij}=\left[\begin{array}[]{ccc}c_{nij1}&\dots&c_{nijM}\end{array}\right]^{\intercal}\in\mathbb{C}^{M}$ represents the image of the $n$th source, and $\mathbf{A}_{i}=\left[\begin{array}[]{ccc}\mathbf{a}_{1i}&\dots&\mathbf{a}_{Ni}\end{array}\right]\in\mathbb{C}^{M\times N}$ is called the mixing matrix. All the signals are assumed to have zero mean.

With the signal model given in (1), the problem of blind source separation (BSS) becomes one of identifying a demixing matrix $\mathbf{D}_{i}$ such that

where $\mathbf{y}_{ij}=\left[\begin{array}[]{ccc}y_{1ij}&\dots&y_{Nij}\end{array}\right]^{\intercal}\in\mathbb{C}^{N}$ denotes an estimate of the source signal $\mathbf{s}_{ij}$, and $\mathbf{D}_{i}=\left[\begin{array}[]{ccc}\mathbf{d}_{1i}&\dots&\mathbf{d}_{Mi}\end{array}\right]\in\mathbb{C}^{N\times M}$. The difficulty of this identifiying depends on many factors, e.g., number of sources, number of sensors, the nature of the source signals, the property of the mixing system. In this work, we focus on the case where the number of souces is equal to the number of microphone sensors, i.e., $M=N$ and assume that the source signals are mutually independent and their distributions are stationary. Given these conditions, the problem of MBSS can be solved using only the second-order statistics [2].

From (1), the covariance matrix of mixtures is $\mathbf{X}_{ij}=\mathbb{E}\left[\mathbf{x}_{ij}\mathbf{x}_{ij}^{\mathsf{H}}\right]$. Using the signal model given in (1), we obtain

Under the assumption of statistical independence among sources, the covariance matrix of sources $\boldsymbol{\Lambda}_{ij}=\mathbb{E}\left[\mathbf{s}_{ij}\mathbf{s}_{ij}^{\mathsf{H}}\right]$ should be a diagonal matrix, which can be represented as $\boldsymbol{\Lambda}_{ij}=\mbox{Diag}\left(\begin{array}[]{cccc}\lambda_{1ij}&\lambda_{2ij}&\cdots&\lambda_{Nij}\end{array}\right)$. Let us stack all the non-zeros parameters in $\boldsymbol{\Lambda}_{ij}$ to form a large matrix, i.e.,

where $\boldsymbol{\lambda}_{n}\in\mathbb{R}^{I\times J}$ is a matrix consisting of all the parameters associated with the $n$th source, and $\boldsymbol{\lambda}$ is a matrix of size $N\times I\times J$, i.e., $\boldsymbol{\lambda}\in\mathbb{R}^{N\times I\times J}$, which encompass all the parameters of the source model,

In ILRMA, the NMF tool is used to decompose the matrix $\boldsymbol{\lambda}_{n}$, $n=1,\cdots,N$, into the following form:

where $\mathbf{T}_{n}$ is a basis matrix of size $I\times K$ with $K$ being the number of bases, and $\mathbf{V}_{n}$ is called the activation matrix, whose dimension is $K\times J$. This decomposition based source modeling is referred to as the NMF source model.

Another way to model the source model parameters is through a rank-one-tensor decomposition of the matrix $\boldsymbol{\lambda}$, leading to the non-negative CPD (NCPD) based source model. In NCPD, the matrix $\boldsymbol{\lambda}$ is expressed as

where

and $\circ$ denotes the outer product.

In this section, we derive the update rules for source model related parameters using the objective function given in (II-A). For $\overset{\circ}{\boldsymbol{\Lambda}}_{ij}$, the the objective function can be expressed as

Since $\mathbf{U}$ is a unitary orthogonal matrix, (24) can be further expressed as

Direction optimization of $\mathcal{L}(\!\overset{\circ}{\boldsymbol{\Lambda}}_{ij}\!)$ with respect to $\overset{\circ}{\boldsymbol{\Lambda}}_{ij}$ is rather difficult. To circumvent this, we introduce the Jensen’s inequality and the tangent line inequality to obtain the following auxiliary function for (III):

where $\alpha_{okij}$ and $\beta_{oij}$ are two auxiliary variables and the equality holds if and only if $\alpha_{okij}=\frac{t_{oik}v_{okj}}{\sum_{k}t_{oik}v_{okj}}$, $\beta_{oij}=\sum_{k}t_{oik}v_{okj}$.

Identifying partial derivatives of $\mathcal{Q}(\overset{\circ}{\boldsymbol{\Lambda}}_{ij})$ with respect to $\partial t_{oik}$ and $\partial v_{okj}$ and forcing the results to zero, we obtain:

Similarly, the maximum likelihood function with respect to $\mathbf{U}$ can be expressed as

Following the previous analysis, one can obtain the auxiliary function for (III) with respect to $\mathbf{G}$:

where $\hat{\alpha}_{onij}$ and $\hat{\beta}_{nij}$ are two auxiliary variables, and the equality satisfies if and only if $\hat{\alpha}_{onij}=\frac{g_{no}\sum_{k}t_{oik}v_{okj}}{\sum_{o}g_{no}\sum_{k}t_{oik}v_{okj}}$, and $\hat{\beta}_{nij}=\sum_{o}g_{no}\sum_{k}t_{oik}v_{okj}$. Identifying the partial derivative of(III) with respect to $g_{no}$ and forcing the result to zero gives the following update rule:

The update rules of the demixing matrix $\mathbf{D}_{i}$ in cILRMA is similar to that in AuxIVA [9], which are as following:

where $\mathbf{O}_{ni}$ denotes the auxiliary variable, and $\mathbf{e}_{n}$ denotes the $n$th column vector of the identity matrix of size $N\times N$.

The paper presented a clustered source model tailored for ILRMA-based MBASS. Leveraging the NBTD technique, this model defines blocks as outer products of vectors (clusters) and matrices for spectral structure modeling, thus providing interpretable latent vectors. By integrating orthogonality constraints, the model ensures independence among source images. Experimental results demonstrated the superiority of the proposed method over its traditional counterparts in anechoic environments.