Options
On the Application of Blind Source Separation Algorithms to Decompose the Mixed Speech Signals in Digital Hearing Aids
Date Issued
2012
Date
2012
Author(s)
Chung, Huai-Chi
Abstract
As result of receiving outside sounds by microphones, and then amplitude of the sounds amplification by amplifier, the fundamental principles of the traditional hearing aid recorded sounds that mixed signals including speech signals and noises; therefore they amplitude not only the sounds, but also the background noises. Many researchers proposed various algorithms for eliminating some of the specific noises frequency. However, those algorithms usually can’t reduce the sophisticated and changeable noises effectively. With the incessant improved development of the technology, the Blind Source Separation (BSS) elevated chance to eliminate noises. Among BSS, the Independent Component Analysis (ICA) is the dominant methods, which combined by Neural network theory and statistics. In this thesis, we performed a series of simulation based on three different kinds of algorithms, those are Information-MaximizationICA(InfomaxICA), FastICA and Joint Approximate Diagonalization of Eigenmatrices (JADE), and also focus on the separation performance of ICA for separating mixed speech signals. Nevertheless, because of the conventional ICA is not well suit for separating convolutive acoustic signals, therefore we conducted simulations under some special conditions to enable ICA to maintain a certain degree of separating performance. According to our simulation results, after speech signals separated by ICA, we can find the contents of the recovered signals are the same with the original ones, and owing to less destruction of the signals so can the characteristics of original retain more. Therefore, it is a good choice to apply on separating mixed speech signals in digital hearing aids.
Subjects
Independent Component Analysis
Blind Source Separation
Information- Maximization ICA
Fast ICA
Joint Approximate Diagonalization of Eigenmatrices
Type
thesis
File(s)
No Thumbnail Available
Name
ntu-101-R98525063-1.pdf
Size
23.54 KB
Format
Adobe PDF
Checksum
(MD5):12bb36965f6cf7575a5ea1789ec0c446