Use this URL to cite or link to this record in EThOS:
Title: Blind source separation : the effects of signal non-stationarity
Author: Alphey, Marcus J. T.
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2002
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This thesis investigates the effect of non-stationarity reduction, in the form of silence removal, on the performance of blind separation and deconvolution techniques for speech signals. An information-maximisation-based system is used for the separation of instantaneously mixed signals, and a decorrelating system for convolutively mixed signals. An introduction to the concepts of adaptive signal processing, blind signal processing and artificial neural networks is presented. A review of approaches to solving the blind signal separation and deconvolution problems is provided. The susceptibility of the information-maximisation approach to signal non-stationarity is discussed and two methods of silence identification and removal are compared and used to pre-process data before blind separation. The "infomax" approach is used to separate instantaneous mixtures, and is also modified to incorporate silence assessment and removal techniques to form an on-line system. Further modifications are made to the algorithm to investigate the effect of alternative update strategies, and these are compared with experimental results from identical modifications to diverse separating algorithms. A performance metric is used to assess the quality of separation achieved. The application of these techniques to convolutively mixed speech signals is also investigated, using the CoB1iSS algorithm. The effectiveness of the application of the silence removal techniques to both the time domain and frequency domain representations of the outputs is tested. While this form of non-stationarity reduction improves the rate of convergence for instantaneous mixtures, it does not cause any significant improvement in separation performance under most of the experimental conditions tested. No significant difference in performance was noted for the separation of convolutive mixtures in either the time or frequency domain.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available