Loading...

Blind Speech Separation using Sparse Component Analysis

Ghasimi, Majid | 2013

700 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 45341 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Babaei-Zadeh, Masoud
  7. Abstract:
  8. Separation of speech signals has many applications. For example, it is used human-machine communications. This problem can be divided into three categories: overdetermined, determined and underdetermined. If the number of mixtures is greater than the number of sources, the problem is called overdetermined; if it is equal to the number of sources, the problem is called determined and if it is less than the number of sources that problem is called underdetermined. This MS thesis studies the undertermined case. Moreover, speech signals are sparse in time-frequency domain, meaning that in each time-frequency point, usually only one source is active. So, for separatig the speech signals,the mixing matrix can be first estimated by a short-time Fourier transform (STFT) and clustering, and then the signals are recovered by sparse processing. In this thesis, some previous speech separation algorithms are first studied and then and a modified algorithm is proposed, which we call MAP-Based Underdetermined Blind Speech Separation (MBUBSS). This algorithm includes three important components, namely estimating the channel, solving permutation indeterminacy and source recovery. For solving the permutation indeterminacy, a new method is proposed and we show that this method has a better performance in comparison to DOA-based methods. In the source recovery stage, we suggest utilizing the ℓ0 norm minimization in order to improve the speed of the algorithm instead of using ℓ1 norm minimization. We show that ℓ0 minimization by Smoothed ℓ0 (SL0) algorithm is 400 times faster than ℓ1 minimization method, while having almost the same quality
  9. Keywords:
  10. Sparse Decomposition ; Speech Separation ; Sparse Component Analysis (SCA)

 Digital Object List

 Bookmark

No TOC