A screening test for speech pathology assessment using objective quality measures

Vocal dysfunctions and pathologies can be devastating to one´s ability to produce speech properly. A novel method for approaching the problem of speech pathology assessment is presented in this paper. The focus is not on detecting or measuring all possible pathologies, but rather on assessing the vocal quality for those cases where the probability of pathology is high. The system is a screening test that uses objective quality measures that examine both excitation and vocal tract characteristics. The five integrated measures are pitch perturbation, amplitude perturbation, a main cepstral peak measure, the log-likelihood measure and an energy-weighted log-likelihood measure. They are evaluated over six speech phoneme classes and their ability to assess the quality of speech is examined. Ultimately, these measures are to be seamlessly integrated into an overall pathology assessment system using a hidden Markov model (HMM) speech recognizer. To demonstrate the ability of the quality measures to probe the multidimensional perceptual quality space, a neural network-based speech pathology detection scheme was established. This system attained an average classification rate of 85.8% for healthy and pathological speech