Voice Disorders Identification Based on Different Feature Reduction Methodologies and Support Vector Machine Meisam Khalil Arjmandi, Mohammd Pooyan, Hojat Mohammadnejad, Mansour Vali Biomedical Engineering Shahed University Tehran, Iran msmarjmandi@gmail.com Abstract— Identification of voice disorders has been a vital role in our life nowadays. Acoustic analysis can be useful tool to diagnose voice disorders as a complementary technique to other medicine methods such as Laryngoscopy and Stroboscopy. In this paper, we scrutinized feature reduction techniques such as principal component analysis (PCA) and linear discriminant analysis (LDA) as feature subset extraction methods and individual feature selection (IFS), forward feature selection (FFS), backward feature selection (BFS) and branch and bound feature selection (BBFS) as feature subset selection procedures. Performance of each method is evaluated by different classifiers. Between feature selection methods, individual feature selection followed by SVM classifier shows the best recognition rate of 91.55% and AUC of 95.80% among these methods. The experimental results demonstrated that highest performance could be achieved by recognition rate of 94.26% and AUC of 97.94% using linear discriminant analysis along with support vector machine as a classifier. Also this mixture has lowest order of computational complexity in comparison with other architectures. Keywords- voice disorders identification; features subset reduction; support vector machine; linear discriminant analysis; features subset selection. I. INTRODUCTION In modern communities, voice disorders problem due to severe daily activities becomes a major issue. Invasive techniques such as Stroboscopy, Laryngoscopy and Endoscopy are employed by physicians to diagnose of voice impairments; especially disorders disrupt vocal cord mechanism [1]. Rate of health of the vocal folds affect quality of the voice. If the vocal folds become inflamed, some growths may develop on them or they become paralyzed and therefore suffer speech production process. The common disorders are vocal fold paralysis, vocal fold edema, adductor spasmodic dysphonia, A-P squeezing, etc. Disorders usually show up in speech signal in the form of acoustic perceptual measures like hoarseness, breathiness and harshness [2]. The speech results from three components of voice production; voiced sound, resonance and articulation [2]. Voiced sound produced by vocal fold vibration then amplified and modified by the vocal tract resonators (the throat, mouth cavity, and nasal passage) and finally vocal tract articulators (the tongue, soft palate, and lips) modify the voiced sound, therefore produce recognizable words [3]. In this context, we are countered with several types of organic voice disorders such as abductor spasmodic dysphonia, A-P squeezing, cancer, arytenoids dislocation, cyst, and etc [4]. Digital processing of voice signal has proved can be used as a non-invasive technique and an objective diagnosis to assess voice disorders in research setting. The aims of this research could be the evaluation of the performance of laryngitis treatment, surveying pharmacological treatment and rehabilitation effects. In general, short-time and long-time methods are applied as two categories for feature extraction from speech signal. In the previous researches, many long-time parameters such as fundamental frequency (F 0 ), jitter, shimmer, harmonics to noise ratio (HNR), pitch perturbation quotient (PPQ), amplitude perturbation quotient (APQ), normalized noise energy (NNE), voice turbulence index (VTI), frequency amplitude tremor (FATR), glottal to noise ratio (GTR) ([2],[5],[6],[7],[8],[9],[10]) are used in order to evaluate voice system health. Through this study the automatic detection of voice impairments is carried out by means of two feature reduction methods and four feature selection techniques. Then efficiency of each method is examined by different classifiers. This paper presents a novel approach to detect the pathological voice sample from normal one. Efficiency of feature extraction methods, as a feature reduction procedure [12], and feature selection techniques are evaluated. This research attempted to investigate different algorithms for disorders classification and finally introduce the efficient structure to improve recognition rate and computational complexity. 22 features, include long-time acoustic parameters developed by the Massachusetts Eye and Ear Infirmary (MEEI) Voice and Speech Labs [11], are used to identification task. Effect of principal component analysis and linear discriminant analysis-as feature extraction methods- [12] and individual feature selection, forward feature selection, backward feature selection and branch & bound feature selection -as feature selection methods- are examined by different classifiers. The details of long-time features are mentioned in [11]. Proceedings of ICEE 2010, May 11-13, 2010 978-1-4244-6761-7/10/$26.00 ©2010 IEEE 45