Robust impaired speech segmentation using neural network mixture model

Abstract

Abstract-This paper presents a signal processing technique for segmenting short speech utterances into unvoiced and voiced sections and identifying points where the spectrum becomes steady. The segmentation process is part of a system for deriving musculoskeletal articulation data from disordered utterances, in order to provide training feedback. The functioning of the signal processing technique has been optimized by selecting the parameters of the model. The optimization has been carried out by testing and comparing multiple Differential Evolution implementations, including a standard one, a memetic one, and a controlled randomized one. Numerical results have also been compared with a famous and efficient swarm intelligence algorithm. For the given problem, Differential Evolution schemes appear to display a very good performance as they can quickly reach a high quality solution. The binomial crossover appears, for the given problem, beneficial with respect to the exponential one. The controlled randomization appears to be the best choice in this case. The overall optimized system proved to segment well the speech utterances and efficiently detect its uninteresting parts.