THE MAPSPACE DENOISING ALGORITHM FOR NOISE ROBUST SPEECH RECOGNITION

1
THE MAP-SPACE DENOISING ALGORITHM FOR NOISE
ROBUST SPEECH RECOGNITION
Stereo-based Piecewise Affinity Compensation for
Environment
ShihHsiang 2006
2
Reference

• K. Daoudi and C. Cerisara, The map-space
  denoising algorithm for noise robust speech
  recognition, in Proc. ASRU, Cancuun, Mexico,
  2005.
• C. Cerisara and K. Daoudi, Evaluation of the
  SPACE denoising Algorithm on Aurora2, in Proc.
  ICASSP, Toulouse, France, 2005.

3
Outline

• Introduction
• The SPACE algorithm
• The MAP-SPACE algorithm
• Experiments
• Conclusions

4
Introduction

• Robustness techniques can be roughly classified
  into two categories
• Signal Processing
• Achieve noise robustness or denoise the signal
• Adaptation Techniques
• Initial canonical models are transformed to
  represent the new environment
• Require relatively a large amount of adaptation
  data
• The need of speech data transcription in an
  unsupervised mode
• In this paper, they proposed an algorithm, called
  MAP-SPACE, which can be seen as an hybrid
  approach between a denoising and an adaptation
  techniques

5
The SPACE algorithm

• The first step of the algorithm is to model the
  noisy speech y by a mixture of I Gaussians
• The second step is to model clean speech x by a
  mixture of I Gaussians
• Assume that the acoustic region modeled by is
  related to the one modeled by by a certain
  transformation

prior
prior
6
The SPACE algorithm (cont.)
Clean GMM
Noisy GMM
Find a transformation through stereo data
Find a transformation through stereo data
Find a transformation through stereo data
7
The SPACE algorithm (cont.)

• Assume the relationship is deterministic and
  affine
• in each acoustic region
• the mapping transformation is
• This mapping is the basis of their denoising
  algorithm, that is, they assume
• then clean feature estimate is given by

8
The SPACE algorithm (cont.)

• In order to estimate the parameters ,
  MMSE is used. The objective function to minimize
  is
• given that the covariances are diagonal
• If and
  then for each I, the objective function to
  minimize is

where
where
9
The SPACE algorithm (cont.)

• The problem is then equivalent to minimize Fi,n
  w.r.t ai,n and bi,n, for each i and n. Let
• Then the problem becomes
• The solution to this problem is given by
• Thus obtain Ai and bi for each i.

10
The MAP-SPACE Algorithm

• The originality of the SPACE algorithm resides in
  its ability to be easily modified to handle new
  environment
• Assume that SPACE has been performed for some
  initial training noisy condition and given test
  observations
• Use these observations in a MAP criterion to
  adapt the initial noisy speech GMM to the new
  environment
• Such adaptation keeps correspondence between the
  initial and new model parameters
• That is, if the adapted noisy speech GMM is
• for each i, the new estimate is then given by

11
The MAP-SPACE Algorithm (cont.)

• For faster implementation, this estimate can be
  approximated by
• MAP-SPACE has two major advantages w.r.t.
  traditional adaptation techniques
• No transcription of adaptation data is required
  and no assumption on noise alteration is made
• The amount of adaptation data require to achieve
  good estimate is much lower than in traditional
  adaptation
• As compared to SPLICE, MAP-SPACE has the major
  advantage of making no assumption on the type of
  noise which corrupts testing data

12
Experiment

• The results are obtained using pseudo-clean
  trained HMM
• The HMMs modeling the acoustic units are trained
  using the denoised features instead of the clean
  ones
• This strategy provides an approximate modeling of
  residual noise
• Experiments have been conducted on the clean part
  of Aurora 2 test set A, which artificially
  corrupted by adding various types of noise at
  different SNRs (from NOISEX)
• A mixture of 8 Gaussians is used to model the
  noisy and clean speech GMMs.
• Adaptation of the noisy GMM us realized using the
  whole noisy test set

13
Experiment (cont.)

• Experiments on white noise with different SNR
• Match scores are the best (but impracticable)
• Clean models give very poor performance 24.7
  and 9.8
• Experiments on different noise types (5dB)
• Training is done on a white noise and testing on
  some different noise types

14
Conclusions

• MAP-SPACE is a combination between an extension
  of SPLICE and traditional adaptation techniques
• MAP-SPACE have shown is robustness to SNR and
  noisy type mismatch
• In this paper, they only want to prove their
  algorithm is efficient to compensation the
  distortion corrupted by various kind of noise

15
Practical Implementation (on AURORA2)

• The pseudo-clean features are obtained by
  following steps
• For each noisy condition of the training corpus,
  a GMM is trained using the Maximum Likelihood
  criterion
• The corresponding clean GMM is trained on the
  corresponding clean sentences using the MMSE
  criterion
• For each testing condition
• First estimate the SNR (Detect the closest
  training environment)
• For each test sentence, the energy is computed on
  a sliding window of 64ms length
• SNR the highest energy / the lowest energy (in
  the window)
• The closest corresponding SNR of the training
  corpus is found (4 noisy GMMs)

16
Practical Implementation (on AURORA2) (cont.)

• In the second step, the four noisy GMMs for these
  four training conditions are compares, and the
  one that maximizes the likelihood is chosen
• The test corpus is then denoised using the
  parameters of this GMM and its corresponding
  clean GMM
• Evaluation on AURORA2 test set A
• The multi-style training always outperforms than
  SPACE
• There is no apparent stability in the SPACE
  behavior when the number of Gaussians varies

17
Joint Modeling of Clean and Noisy Speech
Distribution

• There may exist different reasons that explain
  the bad results of SPACE and MAP-SPACE
• The most important one is the fact that the
  Gaussian correspondence hypothesis is not
  verified
• This meas, that the MMSE criterion is not the
  best way to build such correspondences
• Joint modeling of clean and noisy speech
  distribution
• Modeling P(x,y)

18
Comparison of SPACE and SPACE-JM
Average accuracy over the 4 noises and 5 SNRs (5
dB, 10 dB, 15 dB, 20 dB and clean) of aurora2
test set A
Average accuracy over the 4 noises of aurora2
test set A at SNR0 dB

• SPACE-JM are much more stable than SPACE results
• The results not only suggest that a better
  Gaussian correspondence is achieved by SPACE-JM,
  but also it is robust to SNR change