Optimization and Parallelization of Monaural Source Separation Algorithms in the openBliSSART Toolkit

Abstract

We describe the implementation of monaural audio source separation algorithms in our toolkit openBliSSART (Blind Source Separation for Audio Recognition Tasks). To our knowledge, it provides the first freely available C++ implementation of non-negative matrix factorization (NMF) supporting the Compute Unified Device Architecture (CUDA) for fast parallel processing on graphics processing units (GPUs). Besides integrating parallel processing, openBliSSART introduces several numerical optimizations of commonly used monaural source separation algorithms that reduce both computation time and memory usage. By illustrating a variety of use-cases from audio effects in music processing to speech enhancement and feature extraction, we demonstrate the wide applicability of our application framework for a multiplicity of research and end-user applications. We evaluate the toolkit by benchmark results of the NMF algorithms and discuss the influence of their parameterization on source separation quality and real-time factor. In the result, the GPU parallelization in openBliSSART introduces double-digit speedups with respect to conventional CPU computation, enabling real-time processing on a desktop PC even for high matrix dimensions.

Modelling User Affect and Sentiment in Intelligent User Interfaces

“…However, a broad range of machine learning toolkits such as WEKA [32] offer a broad selection of alternative intelligence algorithms. A range of suitable and necessary and needed tools include, e. g., such for enhancement of the signal of interest as by blind source separation [30].…”

Active-Set Newton Algorithm for Overcomplete Non-Negative Representations of Audio

“…It has been observed [17], [18] that measures such as the generalized Kullback-Leibler (KL) divergence [4] or Itakura-Saito (IS) [16] divergence are more appropriate for quantifying the modeling error of magnitude spectra of audio, in comparison to the Euclidean distance that is used in many other fields of science. In this work we restrict ourselves to the KL divergence, which is defined as…”

“…Source separation [33], [48], the process of estimating the individual source signals that make up a mixture, is one of the most commonly used applications of nonnegative representations. The majority of the existing source separation algorithms based on non-negative representations use the EM algorithm for estimating the model parameters [2]- [7], [15], [17], [18], [34], [38], and only the way the dictionaries are constructed or the atoms are represented differ. ASNA gives solutions identical to the EM algorithm (provided that they are run for enough iterations), and can therefore replace the EM algorithm in source separation systems that use a fixed dictionary, to give a faster convergence.…”

“…A variety of divergence measures based on the family of Bregman divergences or Csiszár divergences [11,Chapter 2], including the L 2 error [15], the Kullback-Leibler (KL) divergence [4] and the Itakura-Saito divergence [16], have been considered for optimization in the context of audio processing. For the analysis of audio data represented using magnitude or power spectra, NMF algorithms that minimize the KL divergence have been found to be particularly effective [4], [17], [18], because these spectra exhibit a large dynamic range and the KL divergence better captures the resulting non-linear relevance of the magnitudes.…”

Semi-supervised non-negative tensor factorisation of modulation spectrograms for monaural speech separation

“…This divergence measure is commonly used in non-negative tensor and matrix factorisation, [17], producing effective performance with NMF based source separation [5], [18].…”