How to analyse and manipulate nonlinear phenomena in voice recordings

We address two research applications in this methodological review: starting from an audio recording, the goal may be to characterize nonlinear phenomena (NLP) at the level of voice production or to test their perceptual effects on listeners. A crucial prerequisite for this work is the ability to detect NLP in acoustic signals, which can then be correlated with biologically relevant information about the caller and with listeners' reaction. NLP are often annotated manually, but this is labour-intensive and not very reliable, although we describe potentially helpful advanced visualization aids such as reassigned spectrograms and phasegrams. Objective acoustic features can also be useful, including general descriptives (harmonics-to-noise ratio, cepstral peak prominence, vocal roughness), statistics derived from nonlinear dynamics (correlation dimension) and NLP-specific measures (depth of modulation and subharmonics). On the perception side, playback studies can greatly benefit from tools for directly manipulating NLP in recordings. Adding frequency jumps, amplitude modulation and subharmonics is relatively straightforward. Creating biphonation, imitating chaos or removing NLP from a recording are more challenging, but feasible with parametric voice synthesis. We describe the most promising algorithms for analysing and manipulating NLP and provide detailed examples with audio files and R code in supplementary material.This article is part of the theme issue 'Nonlinear phenomena in vertebrate vocalizations: mechanisms and communicative functions'.

Post-Stroke Dysarthria Voice Recognition based on Fusion Feature MSA and 1D

Post-stroke Dysarthria (PSD) is one of the common sequelae of stroke. PSD can harm patients' quality of life and, in severe cases, be life-threatening. Most of the existing methods use frequency domain features to recognize the pathological voice, which makes it hard to completely represent the characteristics of pathological voice. Although some results have been achieved, there is still a long way to go for practical applications. Therefore, an improved deep learning-based model is proposed to classify between the pathological voice and the normal voice, using a novel fusion feature (MSA) and an improved 1D ResNet network hybrid bi-directional LSTM with dilated convolution (named 1D DRN-biLSTM). The experimental results show that our fusion features bring greater improvement in pathological speech recognition than the method that only analyzes the MFCC features, and can better synthesize the hidden features that characterize pathological speech. In terms of model structure, the introduction of dilated convolution and LSTM can further improve the performance of the 1D Resnet network, compared to ordinary networks such as CNN and LSTM. The accuracy of this method reaches 82.41% and 100% at the syllable level and speaker level, respectively. Our scheme outperforms other existing methods in terms of feature learning capability and recognition rate, and will help to play an important role in the assessment and diagnosis of PSD in China.

Machine Learning-Based Scoring System to Predict the Risk and Severity of Ataxic Speech Using Different Speech Tasks

The assessment of speech in Cerebellar Ataxia (CA) is time-consuming and requires clinical interpretation. In this study, we introduce a fully automated objective algorithm that uses significant acoustic features from time, spectral, cepstral, and non-linear dynamics present in microphone data obtained from different repeated Consonant-Vowel (C-V) syllable paradigms. The algorithm builds machine-learning models to support a 3-tier diagnostic categorisation for distinguishing Ataxic Speech from healthy speech, rating the severity of Ataxic Speech, and nomogram-based supporting scoring charts for Ataxic Speech diagnosis and severity prediction. The selection of features was accomplished using a combination of mass univariate analysis and elastic net regularization for the binary outcome, while for the ordinal outcome, Spearman's rank-order correlation criterion was employed. The algorithm was developed and evaluated using recordings from 126 participants: 65 individuals with CA and 61 controls (i.e., individuals without ataxia or neurotypical). For Ataxic Speech diagnosis, the reduced feature set yielded an area under the curve (AUC) of 0.97 (95% CI 0.90-1), the sensitivity of 97.43%, specificity of 85.29%, and balanced accuracy of 91.2% in the test dataset. The mean AUC for severity estimation was 0.74 for the test set. The high C-indexes of the prediction nomograms for identifying the presence of Ataxic Speech (0.96) and estimating its severity (0.81) in the test set indicates the efficacy of this algorithm. Decision curve analysis demonstrated the value of incorporating acoustic features from two repeated C-V syllable paradigms. The strong classification ability of the specified speech features supports the framework's usefulness for identifying and monitoring Ataxic Speech.

Intergenerational choral singing to improve communication outcomes in Parkinson's disease: Development of a theoretical framework and an integrated measurement tool

Purpose: This study presented an initial step towards developing the evidence base for intergenerational choral singing as a communication-focussed rehabilitative approach for Parkinson's disease (PD).Method: A theoretical framework was established to conceptualise the rehabilitative effect of intergenerational choral singing on four domains of communication impairments - motor drive, timing mechanism, sensorimotor integration, higher-level cognitive and affective functions - as well as activity/participation, and quality of life. A computer-assisted multidimensional acoustic analysis was developed to objectively assess the targeted domains of communication impairments. Voice Handicap Index and the World Health Organization's Quality of Life assessment-abbreviated version were used to obtain patient-reported outcomes at the activity/participation and quality of life levels. As a proof of concept, a single subject with PD was recruited to participate in 9 weekly 1-h intergenerational choir rehearsals. The subject was assessed before, 1 week post, and 8 weeks post-choir.Result: Notable trends of improvement were observed in multiple domains of communication impairments at 1 week post-choir. Some improvements were maintained at 8 weeks post-choir. Patient-reported outcomes exhibited limited pre-post changes.Conclusion: This study provided the theoretical groundwork and an empirical measurement tool for future validation of intergenerational choral singing as a novel rehabilitation for PD.

Toward Real-World Voice Disorder Classification

OBJECTIVE

Voice disorders significantly compromise individuals' ability to speak in their daily lives. Without early diagnosis and treatment, these disorders may deteriorate drastically. Thus, automatic classification systems at home are desirable for people who are inaccessible to clinical disease assessments. However, the performance of such systems may be weakened due to the constrained resources and domain mismatch between the clinical data and noisy real-world data.

METHODS

This study develops a compact and domain-robust voice disorder classification system to identify the utterances of health, neoplasm, and benign structural diseases. Our proposed system utilizes a feature extractor model composed of factorized convolutional neural networks and subsequently deploys domain adversarial training to reconcile the domain mismatch by extracting domain-invariant features.

RESULTS

The results show that the unweighted average recall in the noisy real-world domain improved by 13% and remained at 80% in the clinic domain with only slight degradation. The domain mismatch was effectively eliminated. Moreover, the proposed system reduced the usage of both memory and computation by over 73.9%.

CONCLUSION

By deploying factorized convolutional neural networks and domain adversarial training, domain-invariant features can be derived for voice disorder classification with limited resources. The promising results confirm that the proposed system can significantly reduce resource consumption and improve classification accuracy by considering the domain mismatch.

SIGNIFICANCE

To the best of our knowledge, this is the first study that jointly considers real-world model compression and noise-robustness issues in voice disorder classification. The proposed system is intended for application to embedded systems with limited resources.

A mobile-assisted voice condition analysis system for Parkinson's disease: assessment of usability conditions

BACKGROUND AND OBJECTIVE

Automatic voice condition analysis systems to detect Parkinson's disease (PD) are generally based on speech data recorded under acoustically controlled conditions and professional supervision. The performance of these approaches in a free-living scenario is unknown. The aim of this research is to investigate the impact of uncontrolled conditions (realistic acoustic environment and lack of supervision) on the performance of automatic PD detection systems based on speech.

METHODS

A mobile-assisted voice condition analysis system is proposed to aid in the detection of PD using speech. The system is based on a server-client architecture. In the server, feature extraction and machine learning algorithms are designed and implemented to discriminate subjects with PD from healthy ones. The Android app allows patients to submit phonations and physicians to check the complete record of every patient. Six different machine learning classifiers are applied to compare their performance on two different speech databases. One of them is an in-house database (UEX database), collected under professional supervision by using the same Android-based smartphone in the same room, whereas the other one is an age, sex and health-status balanced subset of mPower study for PD, which provides real-world data. By applying identical methodology, single-database experiments have been performed on each database, and also cross-database tests. Cross-validation has been applied to assess generalization performance and hypothesis tests have been used to report statistically significant differences.

RESULTS

In the single-database experiments, a best accuracy rate of 0.92 (AUC = 0.98) has been obtained on UEX database, while a considerably lower best accuracy rate of 0.71 (AUC = 0.76) has been achieved using the mPower-based database. The cross-database tests provided very degraded accuracy metrics.

CONCLUSION

The results clearly show the potential of the proposed system as an aid for general practitioners to conduct triage or an additional tool for neurologists to perform diagnosis. However, due to the performance degradation observed using data from mPower study, semi-controlled conditions are encouraged, i.e., voices recorded at home by the patients themselves following a strict recording protocol and control of the information about patients by the medical doctor at charge.

Physics of phonation offset: Towards understanding relative fundamental frequency observations

Relative fundamental frequency (RFF) is a promising assessment technique for vocal pathologies. Herein, we explore the underlying laryngeal factors dictating RFF behaviours during phonation offset. To gain physical insights, we analyze a simple impact oscillator model and follow that with a numerical study using the well-established body-cover model of the vocal folds (VFs). Study of the impact oscillator suggests that the observed decrease in fundamental frequency during offset is due, at least in part, to the increase in the neutral gap between the VFs during abduction and the concomitant decrease in collision forces. Moreover, the impact oscillator elucidates a correlation between sharper drops in RFF and increased stiffness of the VFs, supporting experimental RFF studies. The body-cover model study further emphasizes the correlation between the drops in RFF and collision forces. The numerical analysis also illustrates the sensitivity of RFF to abduction initiation time relative to the phase of the phonation cycle, and the abduction period length. In addition, the numerical simulations display the potential role of the cricothyroid muscle to mitigate the RFF reduction. Last, simplified models of phonotraumatic vocal hyperfunction are explored, demonstrating that the observed sharper drops in RFF are associated with increased pre-offset collision forces.

Nonlinear analysis of scalp EEGs from normal and brain tumour subjects

Measurement of features from the chaos theory or as popularly known, the concept of nonlinear dynamics, as indicatives of several pathological conditions and cognition states using the electroencephalography (EEG) signal is very popular. In this paper, the analysis of scalp EEG signals of normal subjects and brain tumour patients using the nonlinear dynamic features has been presented. The nonlinear dynamic features that represent the dimensional and waveform complexities of the signal being analyzed have been considered. The statistical analysis of the selected nonlinear dynamic features has been presented. The results show that the nonlinear dynamic features significantly discriminate the brain tumour group from the normal group.

Robust Detection of COVID-19 in Cough Sounds: Using Recurrence Dynamics and Variable Markov Model

COVID-19, otherwise known as the coronavirus, has precipitated the world into a pandemic that has infected, as of the time of writing, more than 10 million persons worldwide and caused the death of more than 500,000 persons. Early symptoms of the virus include trouble breathing, fever and fatigue and over 60% of people experience a dry cough. Due to the devastating impact of COVID-19 and the tragic loss of lives, it is of the utmost urgency to develop methods for the early detection of the disease that may help limit its spread as well as aid in the development of targeted solutions. Coughs and other vocal sounds contain pulmonary health information that can be used for diagnostic purposes, and recent studies in chaotic dynamics have shown that nonlinear phenomena exist in vocal signals. The present work investigates the use of symbolic recurrence quantification measures with MFCC features for the automatic detection of COVID-19 in cough sounds of healthy and sick individuals. Our performance evaluation reveals that our symbolic dynamics measures capture the complex dynamics in the vocal sounds and are highly effective at discriminating sick and healthy coughs. We apply our method to sustained vowel 'ah' recordings, and show that our model is robust for the detection of the disease in sustained vowel utterances as well. Furthermore, we introduce a robust novel method of informative undersampling using information rate to deal with the imbalance in our dataset, due to the unavailability of an equal number of sick and healthy recordings. The proposed model achieves a mean classification performance of 97% and 99%, and a mean F 1 -score of 91% and 89% after optimization, for coughs and sustained vowels, respectively.

Performance of Different Acoustic Measures to Discriminate Individuals With and Without Voice Disorders

The goal of this study is to compare and combine different acoustic features in discriminating subjects with and without voice disorders. A database of 484 adult patients participated in the research. All subjects recorded a sustained vowel /Ɛ/ and underwent a laryngoscopic examination of the larynx. From the results of the laryngeal examination performed by a physician and the auditory-perceptual judgment performed by a Speech-Language Pathologist, the subjects were allocated to the group with (n = 52) and without (n = 432) voice disorder. Four types of acoustic features were used: traditional measures, cepstral measures, nonlinear measures, and recurrence quantification measures. Recordings comprised the emission of the vowel /ε/. Quadratic discriminant analysis was used as classifier. Individual features in the context of traditional, cepstral, and recurrence quantification measures achieved an acceptable performance of ≥70%. Combination of measures improved the classifier performance. The best classification result (86.43% accuracy) was obtained by combining traditional linear and recurrence quantification measures. Results shown that Traditional, Cepstral, and recurrence quantification measures are promising features that capture meaningful information about voice production, which provides good classification performances. The findings of this study can be used to develop a computational tool for voice disorders diagnosis and monitoring.

Vocal Acoustic Analysis

Vocal acoustic analysis is becoming a useful tool for the classification and recognition of laryngological pathologies. This technique enables a non-invasive and low-cost assessment of voice disorders, allowing a more efficient, fast, and objective diagnosis. In this work, ANN and SVM were experimented on to classify between dysphonic/control and vocal cord paralysis/control. A vector was made up of 4 jitter parameters, 4 shimmer parameters, and a harmonic to noise ratio (HNR), determined from 3 different vowels at 3 different tones, with a total of 81 features. Variable selection and dimension reduction techniques such as hierarchical clustering, multilinear regression analysis and principal component analysis (PCA) was applied. The classification between dysphonic and control was made with an accuracy of 100% for female and male groups with ANN and SVM. For the classification between vocal cords paralysis and control an accuracy of 78,9% was achieved for female group with SVM, and 81,8% for the male group with ANN.

An Acoustic-Signal-Based Preventive Program for University Lecturers' Vocal Health

INTRODUCTION

Professional activities of university lecturers involve continued and sustained use of the voice, leading in many cases to increased risk of developing voice disorders. Risk identification followed by the fast application of preventive or corrective measures is a key issue in this context.

OBJECTIVE

Define and implement a preventive program for the vocal health of university lecturers by using acoustic features automatically extracted from voice recordings to identify risk groups and manage preventive or corrective actions MATERIAL AND METHODS: A total of 170 subjects, aged between 18 and 65, were recruited at the San Pedro de Alcántara Hospital and at the University of Extremadura in Cáceres (Spain). They formed three groups-one of 25 people suffering from vocal fold nodules, another of 25 healthy people, and the third of 120 university lecturers. Medical history and voice status assessment was performed, and voice recordings were made following a research protocol. A feature extraction, selection, and classification procedure was applied to the voice recordings to provide the best predictors for discriminating between pathological and healthy voices. The model parameters were then used to determine the lecturers' probability of suffering vocal fold nodules or other pathologies with similar dysphonic speech. These probabilities were used to classify the lecturers into three risk groups-low, medium, and high. These groups were taken as the basis to assign the lecturers to a primary, secondary, or tertiary prevention level. Different preventive or corrective actions were applied for each prevention level.

RESULTS

The best set of predictors comprised sample entropy, correlation dimension, pitch period entropy, glottal noise excitation, and sex, achieving an overall accuracy of 92% with a random forest classifier. They all showed statistically significant differences between vocal fold nodules and healthy groups (P < 0.05). Three out of the four best acoustic features were nonlinear, showing the importance of nonlinear dynamics for clinical practice. The model parameters were applied to the predictors of the lecturers so as to assign them to the different risk groups, leading to 60.8% (73 out of 120) of the lecturers in the low-risk group, 29.2% (35 out of 120) in the medium-risk group, and 10% (12 out of 120) in the high-risk group. The prevention levels were assigned on the basis of this classification and the medical history and laryngological evaluation of some specific subjects. A statistically significant association was found between the voice status and the assigned prevention level (P < 0.001), with there being a clear dependence relationship (Cramér's V = 0.630).

CONCLUSION

It is feasible to develop and apply a preventive voice program for university lecturers that is aided by features automatically extracted from voice recordings. As the program progresses, it is expected that the information automatically provided for the assignment to prevention levels will become ever more precise. The method proposed can be extended to other voice professionals and other voice disorders.

Modal and non-modal voice quality classification using acoustic and electroglottographic features

The goal of this study was to investigate the performance of different feature types for voice quality classification using multiple classifiers. The study compared the COVAREP feature set; which included glottal source features, frequency warped cepstrum and harmonic model features; against the mel-frequency cepstral coefficients (MFCCs) computed from the acoustic voice signal, acoustic-based glottal inverse filtered (GIF) waveform, and electroglottographic (EGG) waveform. Our hypothesis was that MFCCs can capture the perceived voice quality from either of these three voice signals. Experiments were carried out on recordings from 28 participants with normal vocal status who were prompted to sustain vowels with modal and non-modal voice qualities. Recordings were rated by an expert listener using the Consensus Auditory-Perceptual Evaluation of Voice (CAPE-V), and the ratings were transformed into a dichotomous label (presence or absence) for the prompted voice qualities of modal voice, breathiness, strain, and roughness. The classification was done using support vector machines, random forests, deep neural networks and Gaussian mixture model classifiers, which were built as speaker independent using a leave-one-speaker-out strategy. The best classification accuracy of 79.97% was achieved for the full COVAREP set. The harmonic model features were the best performing subset, with 78.47% accuracy, and the static+dynamic MFCCs scored at 74.52%. A closer analysis showed that MFCC and dynamic MFCC features were able to classify modal, breathy, and strained voice quality dimensions from the acoustic and GIF waveforms. Reduced classification performance was exhibited by the EGG waveform.

Evaluating the use of neural networks and acoustic measurements to identify laryngeal pathologies

Nineteen acoustical measurements were related to 23 larynx conditions by artificial neural networks (ANNs) and principal component analysis. An exhaustive analysis (combining all possible sets of acoustical measurements as ANN inputs) showed a performance of 99.4% for accuracy and 90.3% for sensitivity and specificity in classifying voice signals into normal and non-normal larynx conditions. In the case of individual larynx condition identification, the general sensitivity drop significantly (6.4%), although some conditions were better identified (including "vocal nodules" and "cysts") then others, reaching 99.8% of sensitivity. We also identified the acoustical measurements that produced the best classification results.

Enhanced Living by Assessing Voice Pathology Using a Co-Occurrence Matrix

A large number of the population around the world suffers from various disabilities. Disabilities affect not only children but also adults of different professions. Smart technology can assist the disabled population and lead to a comfortable life in an enhanced living environment (ELE). In this paper, we propose an effective voice pathology assessment system that works in a smart home framework. The proposed system takes input from various sensors, and processes the acquired voice signals and electroglottography (EGG) signals. Co-occurrence matrices in different directions and neighborhoods from the spectrograms of these signals were obtained. Several features such as energy, entropy, contrast, and homogeneity from these matrices were calculated and fed into a Gaussian mixture model-based classifier. Experiments were performed with a publicly available database, namely, the Saarbrucken voice database. The results demonstrate the feasibility of the proposed system in light of its high accuracy and speed. The proposed system can be extended to assess other disabilities in an ELE.

Voice Disorder Classification Based on Multitaper Mel Frequency Cepstral Coefficients Features

The Mel Frequency Cepstral Coefficients (MFCCs) are widely used in order to extract essential information from a voice signal and became a popular feature extractor used in audio processing. However, MFCC features are usually calculated from a single window (taper) characterized by large variance. This study shows investigations on reducing variance for the classification of two different voice qualities (normal voice and disordered voice) using multitaper MFCC features. We also compare their performance by newly proposed windowing techniques and conventional single-taper technique. The results demonstrate that adapted weighted Thomson multitaper method could distinguish between normal voice and disordered voice better than the results done by the conventional single-taper (Hamming window) technique and two newly proposed windowing methods. The multitaper MFCC features may be helpful in identifying voices at risk for a real pathology that has to be proven later.

A new algorithm for quadratic sample entropy optimization for very short biomedical signals: application to blood pressure records

This paper describes a new method to optimize the computation of the quadratic sample entropy (QSE) metric. The objective is to enhance its segmentation capability between pathological and healthy subjects for short and unevenly sampled biomedical records, like those obtained using ambulatory blood pressure monitoring (ABPM). In ABPM, blood pressure is measured every 20-30 min during 24h while patients undergo normal daily activities. ABPM is indicated for a number of applications such as white-coat, suspected, borderline, or masked hypertension. Hypertension is a very important clinical issue that can lead to serious health implications, and therefore its identification and characterization is of paramount importance. Nonlinear processing of signals by means of entropy calculation algorithms has been used in many medical applications to distinguish among signal classes. However, most of these methods do not perform well if the records are not long enough and/or not uniformly sampled. That is the case for ABPM records. These signals are extremely short and scattered with outliers or missing/resampled data. This is why ABPM Blood pressure signal screening using nonlinear methods is a quite unexplored field. We propose an additional stage for the computation of QSE independently of its parameter r and the input signal length. This enabled us to apply a segmentation process to ABPM records successfully. The experimental dataset consisted of 61 blood pressure data records of control and pathological subjects with only 52 samples per time series. The entropy estimation values obtained led to the segmentation of the two groups, while other standard nonlinear methods failed.

Computational study of effects of tension imbalance on phonation in a three-dimensional tubular larynx model

OBJECTIVES

The present study explores the use of a continuum-based computational model to investigate the effect of left-right tension imbalance on vocal fold (VF) vibrations and glottal aerodynamics, as well as its implication on phonation. The study allows us to gain new insights into the underlying physical mechanism of irregularities induced by VF tension imbalance associated with unilateral cricothyroid muscle paralysis.

METHODS

A three-dimensional simulation of glottal flow and VF dynamics in a tubular laryngeal model with tension imbalance was conducted by using a coupled flow-structure interaction computational model. Tension imbalance was modeled by reducing by 20% the Young's modulus of one of the VFs, while holding VF length constant. Effects of tension imbalance on vibratory characteristic of the VFs and on the time-varying properties of glottal airflow as well as the aerodynamic energy transfer are comprehensively analyzed.

RESULTS AND CONCLUSIONS

The analysis demonstrates that the continuum-based biomechanical model can provide a good description of phonatory dynamics in tension imbalance conditions. It is found that although 20% tension imbalance does not have noticeable effects on the fundamental frequency, it does lead to a larger glottal flow leakage and asymmetric vibrations of the two VFs. A detailed analysis of the energy transfer suggests that the majority of the energy is consumed by the lateral motion of the VFs and the net energy transferred to the softer fold is less than the one transferred to the normal fold.

Automatic detection of pathological voices using complexity measures, noise parameters, and mel-cepstral coefficients

This paper proposes a new approach to improve the amount of information extracted from the speech aiming to increase the accuracy of a system developed for the automatic detection of pathological voices. The paper addresses the discrimination capabilities of 11 features extracted using nonlinear analysis of time series. Two of these features are based on conventional nonlinear statistics (largest Lyapunov exponent and correlation dimension), two are based on recurrence and fractal-scaling analysis, and the remaining are based on different estimations of the entropy. Moreover, this paper uses a strategy based on combining classifiers for fusing the nonlinear analysis with the information provided by classic parameterization approaches found in the literature (noise parameters and mel-frequency cepstral coefficients). The classification was carried out in two steps using, first, a generative and, later, a discriminative approach. Combining both classifiers, the best accuracy obtained is 98.23% ± 0.001.

Quantitative acoustic measurements for characterization of speech and voice disorders in early untreated Parkinson's disease

An assessment of vocal impairment is presented for separating healthy people from persons with early untreated Parkinson's disease (PD). This study's main purpose was to (a) determine whether voice and speech disorder are present from early stages of PD before starting dopaminergic pharmacotherapy, (b) ascertain the specific characteristics of the PD-related vocal impairment, (c) identify PD-related acoustic signatures for the major part of traditional clinically used measurement methods with respect to their automatic assessment, and (d) design new automatic measurement methods of articulation. The varied speech data were collected from 46 Czech native speakers, 23 with PD. Subsequently, 19 representative measurements were pre-selected, and Wald sequential analysis was then applied to assess the efficiency of each measure and the extent of vocal impairment of each subject. It was found that measurement of the fundamental frequency variations applied to two selected tasks was the best method for separating healthy from PD subjects. On the basis of objective acoustic measures, statistical decision-making theory, and validation from practicing speech therapists, it has been demonstrated that 78% of early untreated PD subjects indicate some form of vocal impairment. The speech defects thus uncovered differ individually in various characteristics including phonation, articulation, and prosody.

Nonlinear dynamic modeling of impaired voice

This paper presents a nonlinear dynamic model for the purpose of modeling vowels uttered by patients who have problem in the control of voice box muscles. The proposed model will be utilized in the detection of speech pathologies and also automatic speech recognition systems to enhance patients' communication capabilities. The model of this study utilizes feedback, and also a sigmoid nonlinear function which is not included in the linear speech production models. The nonlinear function allows for the higher order dynamics of the signal to be captured and feedback increases dynamicity of the model. The model of the current research was applied to discriminate between few voice pathologies and normal cases. The statistical analysis of the parameters of the trained model showed that these parameters can provide independent and distinct features with which pathological classes can be discriminated.