LOW-RESOURCE NOISE-ROBUST FEATURE POST-PROCESSING ON AURORA 2.0

1
LOW-RESOURCE NOISE-ROBUST FEATURE POST-PROCESSING
ON AURORA 2.0

• Chia-Ping Chen, Jeff Bilmes and Katrin Kirchhoff
• SSLI Lab
• Department of Electrical Engineering
• University of Washington
• Presenter Shih-Hsiang(??)

ICSLP 2002
2
Introduction

• The performance of ASR systems often decreases
  dramatically when the noise level increases
• The degradation is minor when the signal-to-noise
  ratio (SNR) is high, but quite significant at low
  SNR level
• In the past, a variety of techniques have been
  proposed
• Principle component analysis and a discriminative
  neural network (Ellis et al. 2001)
• Missing Data theory (Cooke et al. 2001)
• Voice activity detector (VAD) and variable frame
  rate are used to drop noisy feature vector to
  reduce insertion error (John et al. 2001)
• Nonlinear spectral subtraction, noise masking,
  feature filters, and model adaptation (Lieb et
  al. 2001)
• data-driven temporal filters, on-line mean and
  variance normalization, voice activity detection,
  and server side discriminate features are
  integrated together to improve noise robustness
  (Morgan et al. 2001)
• etc

3
Literature Review

• Ellis et al. 2001
• John et al. 2001
• Variable frame rate processing
• An observation vector is discarded if it does not
  differ much from the previous observation vector.
  In our implementation of VFR, frame-to-frame
  variation is estimated as the Euclidean norm of
  the sub-vector corresponding to the
  delta-cepstrum.
• Voice activity detection

4
Literature Review

• Morgan et al. (2001)

5
Proposed method

• The first step is standard mean subtraction (MS)
• The second step is variance normalization (VN)
• The third step is auto-regression moving average
  (ARMA)

feature vector (cepstral coefficient)
the order of the ARMA filter
6
Choosing a proper order M of the filter
The transfer function is
The frequency response of the ARMA filter of
order M is
There are zeros in the frequency
response of the ARMA filter is approximately
proportional to its order It support that a
large M will perform poorly since it could filter
out important speech information
7
Gain and phase shifts of the ARMA filter
8
The time sequences of the cepstral coefficient c1
for the digit string 5376869 corrupted with
different levels of noises
9
Evaluation

• Evaluate on Aurora 2.0 noisy digits database
• Two training sets and three test sets
• Training sets clean training set only /
  multi-condition speech
• Test sets stationary-noise sets /
  non-stationary-noise sets / convolutional noise
• 7 different levels of noises
• Clean, 20dB, 15 dB, 10dB, 5dB, 0dB, -5dB
• Recognizer
• Simple HMM-based system using whole-word models
• Zero Nine and Oh 16 states per word, 3
  mixture Gaussian per state
• silence 3-states

10
Recognition results
Word accuracies (as percentages)
Top multi-condition training Bottom clean
training
11
A comparison of different orders of the ARMA
filtering

• A small M will retain the short-term cepstral
  information but is more vulnerable to noise
• A large M will make the processed features less
  corrupted by noise, but the short-term cepstral
  information will be lost.

Top multi-condition training Bottom clean
training
12
Test the effectiveness of proposed technique

• The results show that while variance
  normalization and mean subtraction improves
  performance over the baseline, the addition of
  the ARMA filter provides significant further
  improvements

13
Comparison of different filter

• causal ARMA filter
• non-causal MA filter
• causal MA filter

14
Comparison of different filter (cont.)
Top multi-condition training Bottom clean
training