Blind speech dereverberation using multiple microphones

1
Blind speech dereverberation using multiple
microphones

• Inseon JANG, Seungjin CHOI
• Intelligent Multimedia Lab
• Department of Computer Science and Engineering,
  POSTECH
• jinsn_at_postech.ac.kr
• Seungjin_at_postech.ac.kr

2
Outline

• Introduction
• What is the Reverberant speech ?
• Previous approaches for Speech dereverberation
• Blind speech dereverberation using multiple
  microphones
• Blind Equalization using multiple microphones
• Single Input Multiple Output (SIMO) system
• Subspace Method
• Deterministic Method
• Results

3
What is the Reverberant Speech ?

• Reverberant speech
• cf) Noisy speech
• The degrading component of the case of
  reverberation is dependent on previous speech
  data,
• whereas the degrading component of the case of
  noise speech is independent of speech.

4
Previous approaches for Speech dereverberation

• Cepstrum based approach
• Adaptive microphone array processing
• Blind Deconvolution
• Temporal envelope filtering
• Multi-Microphone sub-band envelope estimation
• Wavelet transform extrema clustering
• Maximum-kurtosis subband adaptive filtering
• Using LP Residual signal
• Using Probabilistic Models

5
Blind Equalization using multiple microphones
SIMO system (1/2)
6
Blind Equalization using multiple microphones
SIMO system (2/2)

• where is the filtering matrix
• For virtual channel,

7
Subspace Method

• By orthogonality between the noise and the signal
  subspace,
• the column of are orthogonal to any
  vector in the noise
• subspace
• for
• Subspace-Based Parameter Estimation Scheme
• Minimization of the quadratic form

8
Deterministic Method (1/2)

• Cross Relation Approach

9
Deterministic Method (2/2)

• Channel estimate
• Equivalently, the channel estimate can be
  obtained from
• the singular vector of associated with
  the
• smallest singular value

10
Result (1/3)Reverberant signal and Dereverberant
signal
11
Result (2/3)Dereverberation using Subspace method

• Channel length 654
• Test size 5000
• Result
• MSE 1.3608e-007

12
Result (3/3)Dereverberation using Deterministic
method

• Channel length 654
• Test size 1000
• Result
• MSE 7.7074e-018