Dissertation > Excellent graduate degree dissertation topics show

Research of Speech Separation Based on Independent Component Analysis

Author: LiuPeng
Tutor: ZhouJun
School: Shanghai Jiaotong University
Course: Signal and Information Processing
Keywords: Speech separation independent component analysis convolutive mixtures separation frequency domain method ambiguities permutation alignment
CLC: TN912.3
Type: Master's thesis
Year: 2009
Downloads: 130
Quote: 0
Read: Download Dissertation

Abstract


Speech separation has been a hot topic in signal processing society recently years, which has many applications and influence in telephone conference, hearing aid, portable devices, speech recognition. Blind signal processing is a useful method in speech separation, in which the term“blind”means that the source itself and the transmission channel is unknown. Independent component analysis is the theoretical basis of blind signal separation, which can be used in various signal processing fields including communications, image, speech, biology, radar, seismic, sonar and etc. The thesis starts with the research of basic theories in BSS and then goes to speech separation algorithms based on convolutive model in frequency domain. The mainly works are following.There are many studies in instantaneous mixing model in time and real domain, therefore in real world speech signals are convolved and will be transformed into complex value in frequency domain method. We study algorithms using the complex information to separate complex signals and the permutation problem in frequency domain method of convolution model, propose a new scheme based on separation matrix initialization, and have a performance comparison later.The main method of convolutive speech separation in frequency domain is that transforming the convolution in time domain to multiple computations. So the complex time domain convolutive model becomes frequency-domain instantaneous models in each bin. And we can use the well developed instantaneous algorithms to estimate separating matrices. Therefore after the transformation back to time domain, we can get the deconvolution FIR filters and have the estimated sources. But this frequency domain method brings in the permutation problem which causes performance decrease in separation. Let us assume for a couple frequency-domain signals after the separating in each frequency bin, we cannot assure that each output channel only consists of the components from the same source. So if we simply transform the separating matrices into time domain, this will cause the deconvolution failure.Solving the permutation problem can be called alignment. Traditional methods such as using mutual parameters and geometric DOA(direct of arrival) can be influenced by permutation in last frequency bin and the first step separation, therefore have not shown good results in robust and accuracy in the scheme that isolating separating step with permutation alignment. The thesis studies permutation alignment algorithms using correlation coefficient and DOA estimation, implements the DOA based method. We also have a study on the performance of DOA method and its influence on the final separating results. We explore the fundamental limitations of the above alignment algorithms and then propose a new scheme based on separation matrix initialization, which is also a geometric approach, consider the separating step together with alignment, can solve the permutation when separation. Because it has better results in robust and accuracy, the final estimated sources are better in both objective and subjective evaluation targets.Permutation problem remains being tough and still needs better solutions. The thesis also introduces a new frequency domain blind separation scheme using frequencies dependency, which can in theory probably avoid the permutation.

Related Dissertations

  1. The Research of Information Hiding Algorithm Based on Noise Visibility Function and the Independent Component Analysis,TP391.41
  2. Complex environment speech enhancement method,TN912.35
  3. Speech Separation Based on Computational Auditory Scene Analysis,TP391.42
  4. ICA based on redundant wavelet and image watermarking method,TP309.7
  5. Algorithm and Application of Blind Source Separation Based on an Improved Particle Swarm Optimization,TN911.7
  6. Underdetermined Anechoic Blind Speech Separation Based on Time-Frequency Sparsity,TN912.3
  7. Blind Separation of Multiple Speech Signals,TN912.3
  8. Research on Speech Separation Based on Spatial Filtering,TN912.3
  9. Research on Speech Separation Technology of Mobile Robot Hearing for Navigation System,TN912.3
  10. The Research of Speech Segmentation Based on Oscillatory Neural Network,TN912.3
  11. Study on Signal Processing of Distributed Optical Fiber Pre-warning System,TE973.6
  12. Feature Extraction and Analysis of High-resolution SAR Images for Target Classification,TN957.52
  13. Research of Blind Signal Processing Technology in the RFID System,TP391.44
  14. Study on Vibration Evaluation and Faults Feature Extraction of Marine Power Equipment,U672
  15. Research on Underdetermined Convolutive Speech Signal Separation Methods,TN912.3
  16. Fault Detection and Diagnosis of Rotating Machinery,TH165.3
  17. A Study of Lung Sound Extraction Based on Blind Source Separation,TN911.72
  18. Paradigm Design and Algorithm Research for P300-based BCI,TP334.7
  19. The Several Trials Extraction of the Sptrague-Dawley Rats’ Flash Visual Evoked Potentials and the Preliminary Research about the Action Potentials,Q42
  20. Measurement of Chlorophyll Content and Distribution in Cucumber Leaves Using Hyper-spectral Imaging Technique,TP391.41
  21. Based on independent component analysis database watermarking technology research,TP309.7

CLC: > Industrial Technology > Radio electronics, telecommunications technology > Communicate > Electro-acoustic technology and speech signal processing > Speech Signal Processing
© 2012 www.DissertationTopic.Net  Mobile