Nonnegative Matrix Factorization (NMF) is an approximative low-rank matrix factorization which is frequently applied for source separation of audio signals (see e.g. [1]). The quality of source separation algorithms using NMF strongly depends on the initialization of the NMF. Very often, random values are used for initialization. Several other initialization strategies have been developed, with the aim to find better initial estimates, thus leading to a better resulting factorization. Most of these deterministic initialization methods use singular value decomposition (SVD). In this paper we introduce a new initialization scheme for audio source separation, based on complex SVD. We also evaluate several different state-ofthe-art initializations in an audio source separation environment. We analyze the effect of the different methods on different kinds of mixtures and show, that our simple but efficient method leads to better results than other SVD-based initializations.
[1]
Christos Boutsidis,et al.
SVD based initialization: A head start for nonnegative matrix factorization
,
2008,
Pattern Recognit..
[2]
Reza Boostani,et al.
An Efficient Initialization Method for Nonnegative Matrix Factorization
,
2011
.
[3]
Sirish L. Shah,et al.
Non-Negative Matrix Factorization for Detection and Diagnosis of Plantwide Oscillations
,
2007
.
[4]
Roland Badeau,et al.
Blind Signal Decompositions for Automatic Transcription of Polyphonic Music: NMF and K-SVD on the Benchmark
,
2007,
2007 IEEE International Conference on Acoustics, Speech and Signal Processing - ICASSP '07.
[5]
C. D. Meyer,et al.
Initializations for the Nonnegative Matrix Factorization
,
2006
.