Dissertation > Excellent graduate degree dissertation topics show

Research on Underdetermined Convolutive Speech Signal Separation Methods

Author: LiuBoQuan
Tutor: ZengYiCheng
School: Xiangtan University
Course: Microelectronics and Solid State Electronics
Keywords: Blind Speech Separation Nonlinear binary frequency masking Non - negative matrix factorization Fast relative Newton method
CLC: TN912.3
Type: Master's thesis
Year: 2010
Downloads: 88
Quote: 0
Read: Download Dissertation

Abstract


Voice source signal in the field of speech signal processing, mixed voice signal from speaker isolated, become a hot research topic and difficult, important research direction in the speech signal processing, speech recognition, speech enhancement and so on has a very active role. Blind source separation (Blind source separation, BSS), refers to the unknown time of the source signal and the transmission channel parameters, only by the observed signal obtained source signal process. Most of today's study, voice blind source separation algorithm the number of strict requirements of observed signals is greater than or equal to the number of source signals. However, in the actual situation, it is often underdetermined, i.e. the number of signals of the observer will be less than the number of the case of the source signal. Actual situation must also take into account the impact of environmental factors on the signal propagation will produce delayed effects, mathematical representation is in the form of convolution. Therefore, the search for efficient underdetermined blind speech separation of convolutive mixtures has very great practical significance and value. In this thesis, blind source separation method for underdetermined convolutive mixed voice study: (1) based on fast independent component analysis and adaptive nonlinear binary frequency masking blind speech separation method. Mixed voice signal input fast independent component analysis, the results adaptive nonlinear binary frequency masking; repeated this two-step process, until you isolate all voice source signal. The separated voice source signal, and then by binary frequency masking merger can improve the quality of the output, the separated speech signal is still able to retain the effect of the two-channel stereo. The experiments show that the performance of the method is much better than the the DUET method, and BLUES methods, a substantial increase in signal-to-noise ratio gain. (2) blind speech separation methods based on non-negative matrix factorization (NMF). This method uses the Gaussian component source signal short-time Fourier transform (STFT) said Gaussian component based Itakura - Saito (Itakura-Saito (IS)) factor of the divergence of the non-negative matrix factorization. The great expectations (EM) algorithm for solving parameter and signal restructuring. The method is applied to two-channel stereo signal the blind separation experiments, the experimental results show the effectiveness of the method. (3) based on the fast relative Newton method and smoothing techniques multiplier blind speech separation. The use of the sparsity of the voice signals and voice signals between independent features, use fast relative Newton method makes Newton method, seek the Heisen array step greatly simplified, greatly improving the speed of operation. Multiplier smoothing techniques applied to the largest type of Lagrange multiplier function smooth approximation, to obtain an extension of the augmented Lagrangian method. This method ensures the rapid convergence of the smoothing factor without increasing the dimension of the problem, and achieved good separation. In this paper, three underdetermined blind speech separation of convolutive mixtures has some theoretical significance and application value.

Related Dissertations

  1. Research on Co-clustering and Application,TP311.13
  2. Non-negative Local Coordinate Factorization for Image Representation,TP391.41
  3. Non-negative matrix factorization based on sparse image retrieval,TP391.41
  4. Research on Human Abnormal Behavior Detection and Recognition in Intelligent Video Surveillance,TP391.41
  5. Internet Oriented Sentiment Orientation Analysis of Chinese Public Opinion,TP391.1
  6. Hyperspectral Unmixing Based on Nonnegative Matrix Factorization,TP751.1
  7. Research on Hyperspectral Remote Sensing Image Fusion and Quality Evaluation Method,TP751
  8. Spam Filtering System Study Based on Non-negative Matrix Factorization,TP393.098
  9. Research of Single-document Summarization Based on Semantics,TP391.1
  10. The Research on Image Protection Technology Based on Perceptual Hashing,TP391.41
  11. Underdetermined Anechoic Blind Speech Separation Based on Time-Frequency Sparsity,TN912.3
  12. A Non-negative Matrix Factorization Model with Sparse Constraint and Its Initialization,O151.21
  13. Support vector machine and method for non-negative matrix factorization theory application,TP391.41
  14. Constraint -based sparse NMF algorithm for face recognition Aging Simulation and its application to face recognition,TP391.41
  15. Research on Space-Borne Multi-Sensor Image Fusion Based on Non-Negative Matrix Factorization,TP751
  16. Face Recognition Method Research and Implementation,TP391.41
  17. Research on Chinese Text Categorization Based on NLP and Non-negative Matrix Factorization,TP391.1
  18. Study on Algorithm of Image Base Learning Using Non-Negative Sparse Coding,TP301.6
  19. Non-negative Matrix Factorization and Its Applications to Face Recognition,TP391.41
  20. Nonnegative Matrix Factorization with Sparseness Constraint,O151.21

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