Hidden Markov models and multilayer perceptrons in a practical speech recognition application using the TMS320C50

In this report the development of a speech recognising telephone is described The speech recognition unit is designed to handle 13 words: the digits 0 to 9 and 3 commands. The TMS32OC5O, a Digital Signal Processor from Texas Instruments, is the core of the system. It is described briefly, together with the usefulness for speech recognition applications. FFT-based pre-processing prepares for two techniques that are applied to perform the recognition task: the Iziltilaver Perceptron and the Hidden Markov Models Both methods are evaluated on basis of usefulness. development time, and performance. For the Hidden Markov Model class/ier a tool is implemented. Generation of observation sequences is based on L VQ using a Kohonen Network. The itIultilayer Perceptron and the Hidden Markov Model Class/ler takes nearly the same amount of development time. The Multilayer Perceptron performs best on the TMS32OC5O, it requires less memory and processor time. The Multilayer Perceptron appears to be the best method in this application, though the Hidden 1tlarkor it/ode! class qfler is realised ii'ith more ease.