LOG-ENERGY DYNAMIC RANGE NORMALIZATON FOR ROBUST SPEECH RECOGNITION

1
LOG-ENERGY DYNAMIC RANGE NORMALIZATONFOR ROBUST
SPEECH RECOGNITION

• Weizhong Zhu and Douglas OShaughnessy
• INRS-EMT, University of Quebec
• Montreal, Quebec, H5A 1K6, Canada
• Presenter Chen, Hung-Bin

ICASSP 2005
2
Outline

• Introduction
• Observation
• Energy dynamic range normalization method1
• Energy dynamic range normalization method2
• Experiment
• Conclusion

3
Introduction

• Automatic Speech Recognition (ASR) has been in
  commercial application for decades but still has
  severe limitations.
• Accuracy of speech recognition degrades rapidly
  when speech is distorted by noise.
• Methods to overcome the effects of noise must be
  applied in order to achieve good recognition
  accuracy in real speech recognition applications
  where various types of noises may exist.
• Robust speech recognition is one of the most
  challenging areas of speech recognition.

4
Introduction (cont.)

• Methods of robust speech recognition can be
  classified into two approaches.
• front-end processing method is to suppress the
  noise and get more robust parameters
• back-end processing is to compensate for noise
  and adapt the parameters inside the HMM system
• In this paper, we focus on the first approach.
• We try to find a more effective way to remove the
  effects of additive noise for the log-energy
  feature.
• We propose a log-energy dynamic range
  normalization (ERN) method to minimize mismatch
  between training and testing data.
• The dynamic range of log-energy feature sequences
  of an utterance is normalized to a target dynamic
  range.

5
Observation

• Comparing with the log-energy feature sequence
• noisy speech with a 10 dB SNR ratio and that of
  clean speech
• 1. Elevated minimum value,
• 2. Valleys are buried by additive noise energy,
  while perks are not affected as much.

6
Energy dynamic range normalization

• The larger difference on valleys leads to a
  mismatch between the clean and noisy speech.
• To minimize the mismatch,
• we suggest an algorithm to scale the log-energy
  feature sequence of clean speech, in which we
  lift valleys while we keep peaks unchanged.

7
Energy dynamic range normalization

• define a log-energy dynamic range of the sequence
  as follows

8
Energy dynamic range normalization

• Following are the steps of the proposed
  log-energy feature dynamic range normalization
  algorithm

9
Energy dynamic range normalization

• Linear scaling equation may not be the best
  solution.
• We modify the linear scaling equation into
  non-linear scaling equation.

10
Energy dynamic range normalization

• Figure 2 shows a schematic representation of the
  scaling effect of the proposed algorithm.
• The scaling effect is decreased as its own value
  goes up and the maximum of the sequence is
  unchanged.

11
Experiment

• The proposed method was evaluated on the Aurora 2
  database.
• All recognition tests were conducted using the
  HTK recognition toolkit with the setting defined
  for evaluation.
• Speech models are eleven whole word HMMs fixed to
  16 states with 3 diagonal Gaussian mixtures per
  state.
• Two silence models are defined.
• Data in Test A are added to by noises of Subway,
  Babble, Car and Exhibition.
• Data in Test B are added to noises of Restaurant,
  Street, Airport and Station.
• In Test C, besides the additive noise, channel
  distortion is also included.

12
Recognition results

• The results in this section are defined in terms
  of relative improvement (R.I.)
• where NewScore, Baseline are recognition
  accuracies for each test using proposed and
  reference algorithms,

13
Experiment

• Results of table 1 show relative improvements
  with the different target log-energy dynamic
  range.

14
Experiment

• The results of relative improvement in different
  target dynamic ranges using this non-linear
  normalization method are shown in Table 2.
• It achieves a 30.83 highest overall relative
  improvement when the target range is set to 14
  dB.

15
Experiment

• Performance comparisons between linear and
  nonlinear normalization methods for average
  relative improvement at different SNR levels are
  shown in table 3.
• The mean recognition accuracy for each test set
  is obtained by taking the average of the
  recognition accuracies measured in 20, 15, 10, 5
  and 0 dB SNR.

16
Experiment

• Experiment 2 in Table 4
• Here in experiment 2, we answer the questions
• (1) what are the results of techniques like
  cepstral mean and variance normalizations?
• (2) Can the proposed algorithms combine with
  these techniques get an even better result?
• CMN refers to cepstral mean normalization
• process with all 13 parameters
• CVN for cepstral variance normalization
• process with all 13 parameters
• ERN(L) for proposed methods is linear
  respectively
• ERN(N) for proposed methods is non-linear
  respectively

17
Experiment
18
Conclusion

• A log-energy dynamic range normalization
  technique is introduced to improve ASR
  performance in noisy conditions.
• Reducing mismatch in log-energy leads to a large
  recognition improvement.
• It is also confirmed that the proposed algorithm
  can be combined with the cepstral mean or
  variance normalization techniques to achieve an
  even better result.