Design and implementation of a FastICA-based time-frequency algorithm for blind separation of mixed audio signals in a real environment

• Main Author: Pascual, Ronald M.
• Format: text
• Language: English
• Published: Animo Repository 2004
• Subjects: Algorithms; Signal processing; Real-time data processing; Engineering design; Signal Processing
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/3249; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/10087/viewcontent/TG03813_F_Partial.pdf
• Institution: De La Salle University

Blind Signal Separation (BSS) and Independent Component Analysis (ICA) are fairly recent concepts in the field of signal processing that are gaining attention from researchers today because of their various promising applications. Some ICA algorithms, such as the FastICA, have already been developed and validated a few years ago through computer simulations. However, implementing these algorithms on blind separation of mixed audio signals acquired in a real-environment still remains a challenge and poses many other problems. One of the biggest problems with real-environment audio signal separation is the presence of sound propagation delays and reflections (i.e., convolutive mixing). Another thing that needs to be addressed is the problem of separating unsynchronized mixed audio signals acquired by separate acquisition systems. Moreover, one of the greatest challenges that this study dealt with, is regarding to how the basic FastICA algorithm can be modified to work faster especially for the case of lengthy signals. Other challenges include establishing some sets of evaluation of the performance of FastICA in order to determine some of its limitations. This is very important for an evolving field of study like ICA in that future studies may benefit from these findings. The study presented herewith involves design and implementation of a FastICA based time-frequency domain separation algorithm (TFD-FICA) for blind separation of mixed audio signals acquired in a real-environment. Separation in time-frequency domain here is the solution considered in order to improve the capability of FastICA to do BSS for convolutive mixtures. Theoretical validation of the TFD-FICA algorithm showed an average separation improvement of at least 7.27 dB over FastICA. Using best results obtained from the real-environment experiments, a listening-test was also conducted in order to evaluate the separation performance of TFD-FICA. The survey result revealed that 89.5% of the respondents were convinced that the mixed audio signals were separated to at least 60%. A correlation-based algorithm to address the problem of synchronization of the acquired data is also developed and implemented in this study. Performance validations demonstrated that the synchronization method was able to synchronize the mixed signals having acquisition delays lower than 16.6% of the signal length. Furthermore, a modified FastICA version, which is frame-based, is presented here. The said frame-based FastICA showed that the computational efficiency of the original FastICA can be improved by 48.67% on the average, with a minimal decrease in signal-tonoise ratio. Design and Implementation of TFD-FICA 2 Finally, additional computer-simulations experiments were performed in order to evaluate the performance and some limitations of FastICA and TFD-FICA. The first set of simulations, which is a comparative assessment of the efficiencies of the four nonlinearity functions (namely tanh, pow3, gauss and skew), revealed that pow3 works best overall while gauss works the poorest. The second set of simulations showed that TFD-FICA works satisfactorily for signals having propagation delays below 68 milliseconds. Another set of simulations demonstrated that TFD-FICA works satisfactorily for signals having mixing depths below 0.7 or 70%.