Single Mixture Separation of Speech and Interfering Audio Signals Using Subspace Decomposition

This paper presents a method of separating speech and interference signal from their single mixture. The system is based on deriving some independent basis vectors from the mixture spectrogram and clustering them to produce the individual source subspaces. Principal component analysis (PCA) is used to derive some basis vectors reducing the dimension of the mixture spectrogram and independent component analysis (ICA) is used to make the basis vectors independent in their own domain. The independent basis vectors are then grouped into two sets (speech and interfering audio) by employing Kullback-Leibler divergence (KLd) based k-means clustering. Each group of basis vectors is used to decompose the mixture spectrogram into the individual source spectrograms and the time domain source signals are re-synthesized by applying some inverse transformations. The experimental results are noticeable in separating speech and its interfering audio signals.