A Comparative Study of the Multi-Layer Perceptron, the Multi-Output Layer Perceptron, the Time-Delay Neural Network and the Kohonen Self-Organising Map in an Automatic Speech Recognition Task

This paper describes a study of the use of four different neural network techniques for automatic speech recognition (ASR) using two common, real-world application databases. The neural network techniques investigated were the Multi-Layer Perceptron (MLP), the Multi-Output-Layer Perceptron (MOLP), which is an improved version of the MLP , the Time-Delay Neural Network (TDNN) and the Kohohnen Self-Organising Map (SOM). The speech test data consisted of a clean database, acquired in a relatively noise-free room environment, and a telephone database, acquired over conventional dial-up lines. Each database comprised 20 repetitions of 12 isolated words (the digits 0 9 plus ‘nought’ and ‘oh’) each spoken by 25 talkers. Each word was parameterised into a time sequence of 15 frames of an 18-dimension feature vector, consisting of 8 Melfrequency Cepstral Coefficients (MFCCs), the corresponding frame-to-frame MFCC differential coefficients and absolute and differential signal energy coefficients. In a speaker-independent, isolated-word speech recognition task, the respective recognition scores for the MLP, MOLP, TDNN and SOM were 93.2%, 95.5%, 95.1%, and 97.1% respectively for the clean speech database , and 76.3%, 90.5%, 90.5% and 96.8% respectively for the telephone database.