Further Experiments in Variable Frame Rate Analysis for Speech Recognition

摘要

This memo describes the application of a simple VFR analysis to the RSRE AirborneReconnaissance Mission (ARM) system. This is a continuous speech recognition system based on phone-level hidden Markov models (HMMs) which has been developed at the RSRE Speech Research Unit. This section will briefly describe the nature of the data, what VFR analysis is, and its application to automatic speech recognition. Assume that at any 'instant' in time the speech signal can be represented by an ordered set of numbers, or feature vector. This 'instant' is assumed to be short enough that the properties of the speech signal do not change significantly. Any utterance, or collection of words, can then be described as a succession of feature vectors (sometimes referred to as frames). There are areas where the speech signal is relatively constant and hence successive feature vectors will be very similar. In other areas the signal may change rapidly and hence successive feature vectors will be different. In order to reduce the processing time, one obvious solution is to reduce the data (frame) rate. However, parts of the signal which are changing rapidly contain valuable information and so need to be retained. For this reason it is necessary to employ some method of data reduction which actually depends on the data. Variable frame rate coding is such a technique. (JES)