Detection of vocal disorders based on phase space parameters and Lyapunov spectrum

• Based on chaos theory two new sets of features are introduced for distinguishing between normal and pathological voices.
• Mathematical and statistical analyses of the proposed features are conducted.
• Genetic algorithm and linear discriminant analysis are employed to improve performance of the system.

Previous studies have shown that the underlying process of speech generation exhibits nonlinear characteristics. Since linear features cannot represent a nonlinear system thoroughly, this paper employs new sets of non-linear measurement for assessing the quality of recorded voices. Such measurement could be exploited for implementing efficient and convenient systems for diagnosing laryngeal diseases without using invasive methods. Three sets of features based on mutual information, false neighbor fraction, and Lyapunov spectrum are investigated to this end. Furthermore, distributions of the proposed features and their discriminative property are investigated. Moreover, the described procedure benefits from the synergy between different concepts of pattern recognition. First, a genetic algorithm (GA) is invoked to find a-near optimum subset of features. Second, linear discriminant analysis (LDA) is applied to remove remaining redundancies and correlations between selected features. Finally, support vector machine (SVM) is employed for learning decision boundaries. Sensitivity and specificity of 99.3% and 94% respectively were achieved in the simulation results.

Figure optionsDownload as PowerPoint slide