This PhD thesis presents a proof of the contrast function for bounded component analysis (BCA), and discusses one of the limitations of BCA. Additionally, this thesis proposes a blind source separation (BSS) algorithm, inspired by BCA and secondorder statistics (SOS). For unsupervised monaural blind source separation (UMBSS), perceptual features are incorporated into nonnegative matrix factorization (NMF) to establish a family of algorithms. Moreover, a speech production model is combined with NMF to form a Tri-lSNMF algorithm, which decomposes the audio sources into three factors and enforces a harmonic structure and the continuity constraint on the three factors.
|Qualification||Doctor of Philosophy|
|Award date||23 Jul 2020|
|Publication status||Unpublished - 2020|