Pattern recognition and feature selection for the development of a new artificial larynx

A palatometer system is used to read in tongue-palate contact patterns made during speech. The purpose is ultimately to develop an artificial larynx which will operate by determining the intended speech, and then synthesising the voice in a way that will hopefully mimic the user's prelaryngectomy sound. This paper describes the pattern recognition and feature selection techniques used to extract information from the tongue-palate contact patterns, and the effects the various methods have on the classification results. Training and testing datasets were constructed using 50 common words. The following feature extraction methods were used on the datasets: Principal Component Analysis, Fourier Descriptors, Correlation, Image Properties and Generic Fourier Descriptors. Once the features were extracted they were then used as input to a Multi-Layer Perceptron (MLP) Neural Network. The best MLP-based classification rate for the testing dataset was 78%, and this was achieved with the input of Correlation Coefficients. Further research will be conducted to try to improve these classification rates using a voting scheme, and possibly the application of word context.