Chap 10' Environment Robustness

1
Chap 10. Environment Robustness

• Environment difference in Train/Test data
• Channel, speaker normalization - CMN, SBR
• remove the channel effect (in telephone-line,
  microphone speech)
• convolution noise
• Noise compensation PMC
• overcome the additive noise
• Speaker adaptation MLLR
• SA between the SI performance to SD
  performance
• SI ?SA?SD

2
System Modeling of Robust Speech Recognition
System

• Block Diagram
• Acoustic mismatch in train/test data
• Convolution noise (channel effect)
• Additive noise

3
CMN SBR

• CMN(CMS) Cepstral mean normalization
  (Subtraction)
• SBR Signal Bias Removal
• remove the convolution noise
• CMN remove the cepstral mean
• Average over speaker/utterance
• If average over syllable, the information of
  spectrum will be remove.
• 10 error reduction.

4

• Real-time CMN
• Using a LPF to estimate the cepstral mean
• Time constant gt 5 sec
• RASTA, the LPF used is
• SBR remove the cepstral mean
• Ref M. G. Rahim, Biing-Hwang Juang, Signal
  Bias Removal by maximum likelihood estimation for
  robust telephone speech recognition, IEEE Trans.
  on SA, Vol. 4, No. 1, Jan, 1996.

5

• In Cepstral domain
• Assume
• Model
• If

6
PMC parallel model combination

• Training data noisy/clean
• Testing data - noisy
• We cannot train the model with all kinds of SNR,
  types of noise
• (1) Try to remove noise in input signal, Winer
  filter, , etc.
• (2) Try to construct the model from noisy speech
  from speech and noise models PMC (Parallel
  model combination)

7

• Assume the noise is additive, then its also
  additive in freq-domain
• The speech and noise parameters will not in
  addition in MFCC/cepstral/log-spectral domain,
  but in linear-spectral domain
• Noise model ?1-state HMM, noise is stationary
• The parameters of HMM
• model (mean, variance)
• can be added
• in linear-spectral domain?

8

• In linear-spectral domain, the observation prob.
• RV addition ? pdf convolution, for Gaussian pdf
  ? addition of log-likelihood
• Transfer cepstral domain to linear domain
• You can easily find the parameters (mean,
  variance) in linear-domain

9

• Change linear-cepstral to log-spectral (cepstral)
  domain
• In MFCC-domain,
• Problem estimation of SNR

10

• Performance of PMC

11
Maximum Likelihood Linear Regression (MLLR)

• MLLR model adaptation
• Transfer the observation prob. in HMM
• Linear regression
• Mean of mixtures
• Find optimal W using ML criterion
• ? minimize Q-function

12

• Q-function
• Then,
• The observation prob. were cluster into class,
  and multiple linear regressions were used.

13

• Rewrite the above equation
• Finally,

14

• How to find the classes of MLLR
• ? the model with same statistics using same LR
• ? using decision tree to find the classes
• Performance

MLLR
New Model
MAP