Transforms for Speech Recognition

Abstract

In speech processing applications the microphone acts as a transducer to convert sound to an electrical signal which is further converted to a sequence of discrete samples for digital processing. In this raw form the signal does not show readily discernable useful features, and therefore mathematical transformations have been developed to obtain further information that clearly demonstrates characteristics that can be attributed to various types of sounds comprising speech. This is fundamental to the front-end design of all speech recognizers. The importance of an effective and efficient transform for the speech signal is of prime importance since weaknesses at this foundation stage will undoubtedly impair the performance of the following stages. This paper discusses transforms for the speech signal with application to Automatic Speech Recognition. It reviews commonly used representations and modifications made to enhance their performance, and other transforms developed for speech processing and recognition.