Underdetermined Blind Source Separation Combining Tensor Decomposition and Nonnegative Matrix Factorization

Yuan Xie, Kan Xie, Junjie Yang, Shengli XieSymmetry 2018 · DOI

TL;DR

This paper tackles the challenging problem of "underdetermined blind source separation" (UBSS), where we try to disentangle multiple audio sources from fewer microphones than sources. They propose a novel algorithm that combines tensor decomposition for initial mixing matrix estimation with Nonnegative Matrix Factorization (NMF) for source modeling, leading to improved separation performance, especially in reverberant environments.

What Problem Does This Paper Address?

Imagine you're in a room with several people talking, but you only have one microphone. Your goal is to isolate each person's voice. This is the essence of "blind source separation" (BSS). When you have fewer microphones than sources (e.g., 2 microphones, 3 speakers), it becomes an "underdetermined" BSS problem, which is significantly harder. Adding to the complexity, real-world audio signals are often "convolutive," meaning they are affected by echoes and reverberation, making the mixing process more intricate than a simple linear combination. Existing methods often struggle with parameter initialization, leading to poor separation, especially in these challenging underdetermined and convolutive scenarios.

Methods

The core of this paper lies in a novel optimization algorithm designed to overcome the initialization sensitivity of traditional BSS methods. Here's a breakdown:

  • Tensor Decomposition for Mixing Matrix Estimation: Instead of directly estimating the mixing matrix (which describes how sources combine), the authors first use tensor decomposition. This technique leverages the higher-order statistical properties of the mixed signals to uniquely estimate the mixing matrix, providing a robust starting point.
  • Nonnegative Matrix Factorization (NMF) for Source Modeling: NMF is employed to model the spectral characteristics of individual audio sources. It decomposes the power spectrogram of each source into a "basis matrix" (representing spectral patterns) and an "activation matrix" (representing their temporal presence). This is particularly effective for audio signals, which often have sparse spectral structures.
  • Expectation-Maximization (EM) Algorithm for Joint Optimization: The estimated mixing matrix and NMF source models are then iteratively refined using an EM algorithm. This process jointly updates the mixing matrix, NMF parameters, and noise characteristics to improve the overall separation.
  • Wiener Filtering for Spatial Image Estimation: To reconstruct the individual source signals, Wiener filters are used. These filters leverage the learned parameters to estimate the "spatial images" (frequency-domain representations) of each source, effectively separating them from the mixtures.
  • Inverse Short-Time Fourier Transform (STFT): Finally, the estimated source spectrograms are converted back into the time-domain using the inverse STFT, yielding the separated audio signals.

The paper also distinguishes between two cases: linear instantaneous mixtures (simpler, direct combination) and convolutive mixtures (more complex, involving delays and reverberation). They adapt their tensor decomposition approach for the convolutive case by working in the frequency domain.

Datasets

The authors evaluated their proposed algorithm using a combination of synthetic and real-world audio datasets:

  • Dataset A (Synthetic, Linear Instantaneous): Music mixtures with drums, derived from the SiSEC'08 development dataset.
  • Dataset B (Synthetic, Linear Instantaneous): Speech mixtures, also from the SiSEC'08 development dataset.
  • Dataset C (Real-world, Convolutive): Music data with drums, featuring reverberation and microphone spacing simulating a live recording environment.
  • Dataset D (Synthetic, Convolutive): Speech mixtures with realistic reverberation and microphone spacing.

Key Insights and Ideas

  • Tensor Decomposition as a Robust Initializer: The paper highlights that tensor decomposition can provide a significantly more stable and accurate initial estimate of the mixing matrix compared to traditional methods, which is crucial for the success of subsequent EM-based optimization.
  • NMF's Strength in Audio Source Modeling: The effectiveness of NMF in capturing the spectral sparsity and structure of audio sources is leveraged to improve the source models, leading to better separation.
  • Synergy of Tensor Decomposition and NMF: The combination of these two techniques is the paper's main innovation. Tensor decomposition provides a strong initial guess for the mixing matrix, and NMF provides a powerful model for the sources, which the EM algorithm then refines jointly.
  • Addressing Initialization Sensitivity: The primary motivation for this combined approach is to mitigate the notorious sensitivity of EM algorithms to initial parameter values, a common bottleneck in BSS.

Key Results

The proposed algorithm demonstrated superior performance across various metrics (SDR, ISR, SIR, SAR) compared to several baseline methods, including:

  • Linear Instantaneous Case: The proposed "Tensor-EU," "Tensor-KL," and "Tensor-IS" methods consistently outperformed "lo min," "MU," and "EM" algorithms on both music and speech mixtures. For instance, in the linear music case (Figure 2), the proposed methods achieved SDR values around 25-30 dB, significantly higher than the 10-15 dB of some baselines.
  • Convolutive Case: The algorithm also showed marked improvements in the more challenging convolutive mixture scenarios (Figures 7, 8, 10, 11), outperforming methods like "Bm" and "Full-rank."
  • Qualitative Results: Figure 4 and Figure 9 show that the proposed algorithm can effectively recover the waveforms of the original sources, even in complex mixtures.

While the proposed algorithm achieved better separation quality, it did exhibit longer runtime compared to simpler MU and EM methods, though it was faster than the "Full-rank" algorithm.

Limitations and Future Work

The authors acknowledge several areas for future research:

  • Determining the Number of NMF Components: The optimal number of components for NMF is an open problem, and they suggest exploring automatic order selection methods.
  • STFT Window Length Optimization: The choice of STFT window length impacts performance, and further investigation into optimal settings for different audio characteristics is warranted.
  • Handling Source/Microphone Motion: Adapting the algorithm to dynamic scenarios where sources or microphones move (leading to time-varying convolutive mixtures) is a significant challenge for future work.
  • Reducing Computational Cost: While effective, the algorithm's runtime is higher than some baselines. Developing more efficient methods to reduce this time cost is a key future direction.

Summary updated

BibTeX
@article{xie2018underdetermined,
  author = {Yuan Xie and Kan Xie and Junjie Yang and Shengli Xie},
  title = {{Underdetermined Blind Source Separation Combining Tensor Decomposition and Nonnegative Matrix Factorization}},
  journal = {Symmetry},
  year = {2018},
  doi = {10.3390/sym10100521},
}
Figure