Autoregressive hidden Markov model for applied tasks of vocal fold pathology detection

The recognition and training algorithms for autoregressive hidden Markov models were developed in order to solve the task of vocal fold pathology detection. Three databases were created and used for 3 vocal pathologies detection. During the experiments the proposed vocal tract pathology detection system based on autoregressive hidden Markov models and wide-range AFT mel-spectrum provides very high detection accuracy. INTRODUCTION The analysis of speaker individuality is widely used for different tasks, for example, speaker identification and verification, diagnostics of speech producing organs, secure access [1 ], [2]. It is well known that if there is the presence of vocal fold pathology, significant changes can occur that alter the speech production system, resulting in deterioration of voice quality. Analysing such changes we can make a conclusion about the state of person's vocal fold and compare it with the template health person parameters. Subjective testing made by a physician is not able to detect pathology on the earlier stages, except this such testing strictly depends on the physician experience. There is a possibility to perform such analysis objectively based on more or less nonlinear model. From this point of view systems based on hidden Markov models (HMM) are very attractive. Flexible and powerful mathematical apparatus of these models lets to use them for effective temporal information modelling. Speech signal by its nature has two aspects. Firstly, speech of the person is defined by physical parameters, such as vocal tract length, glottal size and so on. Secondly, the speech producing is impossible without neural control of the articulators, which defines the personal learned abilities such as dialect or regional accents, pronunciation, speed and timing of the articulators. For better analysis of speaker pronunciation peculiarities it is necessary to take into consideration interrelations between close frames of the same phonetic unit and loose important information about acoustical structure of the phoneme. In connection with this we introduce an autoregressive hidden Markov model for the task of vocal tract pathology detection, which is similar to the task speaker identification within the framework of such approach [3]. The character vector used for speech analysis greatly influences the voice analysis performance. In order to provide the high accuracy we have to use apriory knowledge about human speech producing and perception. In particular, the knowledge of main psycho-acoustical principles gives us the possibility to cut information not grasping by human ear. Information about high frequencies in the speech signal is very important also, and it is rather reasonable to use it for vocal fold pathology detection. SPEECH PARAMETERS FOR VOCAL FOLD PATHOLOGY DETECTION Now human ear is the best tool for speech analysis and it is not a bad idea to model it in order to solve different speech analysis tasks. The usage of psycho-acoustical principles lets to form speech parameters that reflect all essential speech features. Psycho-acoustical parameters include absolute threshold of hearing, critical bands, simultaneous and temporal masking [4]. Here we briefly review the psycho-acoustical parameters in order to include them to the speech feature vector estimation process. The absolute threshold of hearing is characterized by the amount of energy needed in a pure tone such that it can be detected by a listener in a noiseless environment. The threshold is well approximated by the nonlinear function ~(!}=3.6{!/lOO)r ~~f +10"(!/100~ (1) MAVEBA 1999, Firenze, Italy 108 Models and Analysis of Vocal Emissions for Biomedical Applications