Blind separation of an over-determined mixture of independent speech sources using output decorrelation

• Main Author: Lapira, Edzel R.
• Format: text
• Language: English
• Published: Animo Repository 2003
• Subjects: Speech processing systems; Signal processing; Detectors; Simulation methods; Electrical and Computer Engineering; Electrical and Electronics
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/3119; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/9957/viewcontent/CDTG003591_F_Redacted.pdf
• Institution: De La Salle University

Blind source separation aims to recover or estimate statistically independent speech source signals that have been mixed. The combiner characteristics and the propagation channel are not known, thus the term blind. There are many solutions to blind source separation. Output decorrelation method as the separation croterion is one of the simplest, yet effective solutions. This paper describes a new model for the output decorrelation method by introducing an extra sensor measurement in the demixing section. The classical 2-source x 2-sensor model is modified to a 2-source x 3-sensor model to improve separation behavior. Derived equations for the 2x3 model show an increase in the complexity of the solution of the demixing coefficients, however, performance tests, both objective and subjective, are utilized to gauge the behavior of both models and show an improved separation behavior by as much as an increase in a rating of 1.5 in a scale of 1 to 5.