This item is available under a Creative Commons License for non-commercial use only
Electrical and electronic engineering
Monophonic sound source separation (SSS) refers to a process that separates out audio signals produced from the individual sound sources in a given acoustic mixture, when the mixture signal is recorded using one microphone or is directly recorded onto one reproduction channel. Many audio applications such as pitch modification and automatic music transcription would benefit from the availability of segregated sound sources from the mixture of audio signals for further processing. Recently, Non-negative matrix factorization (NMF) has found application in monaural audio source separation due to its ability to factorize audio spectrograms into additive part-based basis functions, where the parts typically correspond to individual notes or chords in music. An advantage of NMF is that there can be a single basis function for each note played by a given instrument, thereby capturing changes in timbre with pitch for each instrument or source. However, these basis functions need to be clustered to their respective sources for the reconstruction of the individual source signals. Many clustering methods have been proposed to map the separated signals into sources with considerable success. Recently, to avoid the need of clustering, Shifted NMF (SNMF) was proposed, which assumes that the timbre of a note is constant for all the pitches produced by an instrument. SNMF has two drawbacks. Firstly, the assumption that the timbre of the notes played by an instrument remains constant, is not true in general. Secondly, the SNMF method uses the Constant Q transform (CQT) and the lack of a true inverse of the CQT results in compromising on separation quality of the reconstructed signal. The principal aim of this thesis is to attempt to solve the problem of clustering NMF basis functions. Our first major contribution is the use of SNMF as a method of clustering the basis functions obtained via standard NMF. The proposed SNMF clustering method aims to cluster the frequency basis functions obtained via standard NMF to their respective sources by making use of shift invariance in a log-frequency domain. Further, a minor contribution is made by improving the separation performance of the standard SNMF algorithm (here used directly to separate sources) obtained through the use of an improved inverse CQT. Here, the standard SNMF algorithm finds shift-invariance in a CQ spectrogram, that contain the frequency basis functions, obtained directly from the spectrogram of the audio mixture. Our next contribution is an improvement in the SNMF clustering algorithm through the incorporation of the CQT matrix inside the SNMF model in order to avoid the need of an inverse CQT to reconstruct the clustered NMF basis unctions. Another major contribution deals with the incorporation of a constraint called group sparsity (GS) into the SNMF clustering algorithm at two stages to improve clustering. The effect of the GS is evaluated on various SNMF clustering algorithms proposed in this thesis. Finally, we have introduced a new family of masks to reconstruct the original signal from the clustered basis functions and compared their performance to the generalized Wiener filter masks using three different factorisation-based separation algorithms. We show that better separation performance can be achieved by using the proposed family of masks.
Jaiswal, R. (2013) Non-negative matrix factorization based algorithms to cluster frequency basis functions for monaural sound source separation. Doctoral Thesis, Dublin Institute of Technology. doi:10.21427/D7V894