Speech and Audio Signal Processing: Processing and Perception of Speech and Music, Second Edition

CHAPTER 9

STATISTICAL PATTERN CLASSIFICATION

9.1 INTRODUCTION

Audio signals such as speech or music are produced as a result of many causes. For instance, although a speech signal originates in a vocal apparatus, the received signal is generally also affected by the frequency response of the audio channel, additive noise, and so on. The latter sources of variability are unrelated to the message that is communicated. Much of the development of speech-signal analysis is motivated by the desire to deterministically account for as much of this variance as possible. For example, it would be desirable to reduce the spectral variability among multiple examples of the same word uttered with different amounts of background noise. However, with the use of common representations of speech, signals corresponding to the same linguistic message vary significantly.

For this reason, it is often desirable to model an audio signal such as speech as a random process and then to use statistical tools to analyze it. This general class of approaches has proven to be extremely useful for a range of speech-processing tasks, but most notably for speech recognition. However, before seriously discussing the applications of statistics to speech processing, we must introduce some basic concepts. In this chapter we briefly introduce statistical pattern classification (see [5] for a much more thorough treatment). ...