Section IV:
Special Applications
9
Signal and Feature Compensation Methods for Robust Speech Recognition
Rita Singh, Richard M. Stern and Bhiksha Raj
CONTENTS
Introduction
Speech Recognition as Statistical Pattern Classification
Effect of Noise on Speech Recognition Systems
Compensating for the Effects of Noise
Signal Compensation
Linear Spectral Subtraction
Nonlinear Spectral Subtraction
Wiener Filtering
Feature Compensation
Multivariate Gaussian-Based Cepstral Normalization
Vector Taylor Series Compensation
Codeword-Dependent Cepstral Normalization
Cepstral and Spectral High-Pass Filtering
Discussion of Relative Merits of the Methods
Acknowledgments
References
Introduction
As computing, communication, and other electronic devices become physically smaller and attempt to perform increasingly complicated functions, traditional interfaces such as buttons, keyboards, etc. become difficult to use. Speech is a much more natural and simpler interface for such devices, especially if they are to be remotely operated. Viable technology currently exists for the deployment of speech-enabled devices in controlled environmental conditions. However, as these devices are deployed in increasingly difficult operating conditions that are open to uncontrolled noises and acoustical disturbances, the performance of speech recognition systems degrades greatly. This chapter and Chapter 10 are concerned with the subject of development of techniques that reverse this degradation.
Broadly, techniques that enhance environmental robustness for speech recognition systems can be divided into two categories: techniques that operate on speech signals or the features derived from them prior to the recognition process, and techniques that modify the recognition system to perform optimally on incoming noisy speech signals. In this chapter, we review techniques that modify incoming signals or feature vectors. Techniques that modify the structure or parameters of the speech recognition system are discussed in Chapter 10.
For the benefit of readers with a limited background in speech recognition technologies, we begin by reviewing the formulation of automatic speech recognition as a statistical pattern classification process and by discussing how environmental disturbances adversely affect classifier performance. Later sections describe selected signal and feature compensation techniques in current usage, which were chosen on the basis of their efficiency and generality.
Speech Recognition as Statistical Pattern Classification
Automatic speech recognition systems are pattern classifiers designed to solve a rather specific statistical pattern classification problem. A simple example of a statistical pattern classification problem is that of determining the member of a set of N classes C1, C2,…, CN to which a specific vector Xs belongs, knowing that it does belong to one of the classes.
Let be the known distribution of all vectors belonging to class Ci. Let αi be the fraction of all data points that belong to class Ci; αi is also known as the a priori probability of Ci. It can be shown that if the data vector Xs is assigned to a class according to the following rule1:
| (9.1) |
the expected classification error is minimum. In other words, given an infinitely large set of data points to classify, the total number of misclassified points will be minimum if the classification rule above is followed. Pattern classifiers based on the above rule are known as Bayesian classifiers. If the criterion for classification is other than that of minimum expected classification error, e.g., that of minimizing the expected cost of classification (known as minimum risk classification), where the cost may be any function of the output of the classifier, the actual classification rule can vary from the one given above, but its form will still be very similar.
Speech recognition is the problem of determining the sequence of words that were spoken in an utterance, given the recorded signal for that utterance. We can consider the set of all signals that are instances of a particular word sequence to form the class of signals representing that word sequence. Hence, there is a class of signals associated with every possible sequence of words in a language. Statistical speech recognition can be stated as the problem of determining to which of these classes a given signal belongs. This problem can now be treated as an instance of the Bayesian classification ...