Voice Pathology Detection Using a Machine Learning Classifier Enhanced by Hybrid Nature-Inspired Algorithms

This section presents the methodologies used in the identification of speech pathology focusing on ML and DL techniques. The common voice sample features utilized for voice pathology detection are MFCC [5], Jitter, LPCC [7], glottal signal features [8], shimmer, Wavelet entropy [9] and HNR. For categorization, KNN [6], RF [9], SVM [10], DL [11] and MLP [13] models are preferred.

Martínez et al. [14] built a voice pathology detection system using GMM. Four voice attributes such as MFCC, HNR, NNE and GNER were used as features. This system yielded an accuracy of 67%. El Emary et al. [15] focused on the categorization of voice samples into healthy and pathological samples based on MFCC, jitter, and shimmer. The GMM was employed to identify affected voices in a small subset of the SVD database. This subset consisted of 38 healthy and 63 pathological voice samples. This approach reported an accuracy of 82.37%.

In the study conducted by Souissi and Cherif [16], to categorize speech samples into healthy and pathology, the SVM classifier was used and MFCC was used as a feature, while LDA served as the dimension reduction tool. A total of 40 samples of healthy voices and 70 pathological sounds were obtained from the SVD database to evaluate the model. This approach achieved an accuracy of 86%. Amara et al. [17] reported a high accuracy of 95.5% by applying ML classifier to analyse specific /a/ vowel sound. The dataset used for analysis comprised of 45 pathological and 55 healthy voice samples. The above methods used only a small number of speech samples for evaluation. NBN was employed to identify pathological voice samples using MFCC, jitter, shimmer and F [18]. The results showed a highest classification rate of 90%.

Verde et al. [19] selected four features namely MFCC, HNR, jitter and shimmer for voice pathology detection. These features were calculated for each sample, consisting of 685 healthy and 685 pathological voice samples and subsequently used in four different classifiers to assess their classification rate. With a classification accuracy of 85.77%, the SVM classifier had the greatest performance. Lee [13] presented ML models, MFCC, LPCC, NS, and NK for identification of abnormal voices using MLP and CNN. A total of 518 samples consisted of 259 healthy and 259 pathological were obtained, from the SVD database for experimentation.

Table 1. (a) Overview of recent studies on SVD database (b) Strengths and limitations of existing approaches pertaining to voice pathology detection

Another very recent research work [20] proved the usefulness of merged features to detect pathological voices. The authors combined speech and EGG signal elements to improve accuracy in comparison to prior methods. SVM was trained using features extracted from both speech and Electroglottography (EGG), while AlexNet was employed for further analysis. The results indicated that employing the combined features as input yielded a higher classification rate in comparison to solely utilizing EGG or speech signals. Ksibi et al. [21] created for precise identification of abnormal speech on the SVD database, a deep learning model based on CNN-RNN. Signal attributes, MFCC, ZCR, and RMSE were extracted from the input voice samples and then the features were fed as input to the CNN-RNN to perform classification. This approach attained classification rate of 88.83%. Ding et al. [24] created a model for voice pathology detection using ResNet. In this model, features were computed from voice signals, and then fed to deep connected attention-ResNet to predict voice pathology. Table 1(a) provides a summary of recent works on SVD database.

To synthesize the strengths and limitations of the major existing works in the literature, we present a comparative summary in Table 1(b). This highlights the need for an integrated and optimized approach, which our proposed GWOESCA-SVM aims to address effectively.

4.1 Experimental setup

Using MATLAB 2022a platform on an Intel Core i5 CPU, 2.9GHz, 16GB RAM the entire system is implemented. In this investigation, an automated GWOESCA-SVM system is used to distinguish between healthy and pathological voice samples based on merged features from speech and EGG signals, the parameters are listed below:

GWOESCA Parameters are as follows:

• Population size: 30 search agents
• Maximum iterations: 500
• Exploration coefficient (a), linearly decreased from 2 to 0 over iterations (as per standard GWO design)
• Random coefficients (A and C) are drawn from uniform distribution in [0, 1]
• Sine/Cosine diversity control (D) is adaptively updated to enhance search in ESCA phase
• Fitness function: Classification accuracy (evaluated through 10-fold cross-validation)

SVM Tuning Process – parameters are as follows:

• Kernel used is Radial Basis Function (RBF)
• Hyperparameters optimized using Regularization parameter (C) and Kernel parameter gamma
• Search range: C ∈ [0.01, 100] and gamma ∈ [0.0001, 1]
• These were jointly optimized along with the feature subset selection during the GWOESCA iterations.

It is important to note that no data augmentation techniques were applied during this study. As the proposed work relies on a SVM optimized through feature selection, the model architecture does not require the volume of training data typically associated with deep learning models. The feature optimization performed via GWOESCA mitigates overfitting by selecting highly discriminative features from a moderate-sized dataset.

4.2 Evaluation metrics

To evaluate the classification performance, 10-fold cross-validation is employed, wherein the data is divided into 10 smaller sets. Nine sets are utilized for training the system, leaving the tenth set exclusively for testing purposes. Each set takes turns as the testing set, and the process is repeated ten times. As a result, the system’s performance is evaluated based on the average performance obtained over the 10 iterations. The utilization of cross-fold validation effectively prevents the occurrence of data leakage through the experimental procedure, ensuring the reliability of results. The metrics used for assessment are listed in Table 3. True positive samples are those with pathology, whereas True negative samples are those without pathology.

Table 3. Performance metrics

4.3 Results and discussion

In the developed system, feature selection is conducted to diminish the quantity of features while preserving maximum information within the dataset, where GWOESCA is suggested for selecting pertinent features. The simulation parameters of GWOESCA are as follows: number of search agents is set to 30, maximum number of iterations is set at 500 and fitness function used is classification accuracy. Additionally, GWOESCA is used for tuning the parameters of SVM. In optimized SVM parameters, c is 0.3 and gamma is 0.001. The healthy and pathological voice signals for male and female at vowel /a/ are more clearly seen in Figure 4 by showing a box plot of a few selected characteristics.

Male	Female
tu_pian_6.png	tu_pian_7.png
(a) Spectral centroid
tu_pian_8.png	tu_pian_9.png
(b) MFCC
tu_pian_10.png	tu_pian_11.png
(c) Skewness
tu_pian_12.png	tu_pian_13.png
(d) STE
tu_pian_14.png	tu_pian_15.png
(e) ZCR

Figure 4. Distribution of selected features

Effectiveness of the system is analysed in three cases:

Case 1: Classify the data using speech signal features

Case 2: Categorize the data using EGG signal features and

Case 3: Data classification using combined features (Speech + EGG)

The performance metrics of introduced system (Table 3) are calculated and reported in Table 4. In Table 4, first row consisting of 43-dimensional speech features are extracted from the male and female samples. Second row has 43-dimensional EGG features computed from the male and female samples. Third row features are created by adding speech and EGG signals features and it has 48-dimensions. Features are separately computed for male and female samples.

Table 4. Classification outcomes obtained by different features and GWOESCA-SVM focused on gender and individual features

4.3.1 Analysis based on speech features

For males, the classification accuracy was 81.43%, with specificity, recall, precision, GM and F1-score ranging between 81.60% to 84.64%. For females, the accuracy was slightly lower at 78.82% with similar performance metrics. When considering both genders, the classifier achieved higher performance with an accuracy of 88.82%, indicating a slight improvement across all metrics compared to individual gender-based classification.

4.3.2 Analysis based on EGG features

The classification performance using EGG features was slightly lower compared to speech features. For males, the accuracy was 74.90%, and for females, it was 70.78%, both exhibiting similar trends in other metrics. When considering both genders, the accuracy improved to 84.71%, indicating better performance in gender-agnostic classification compared to individual gender-based classification.

4.3.3 Analysis based on combined features

Combined speech and EGG features resulted in significantly improved classification results. The result analysis indicated that combining both speech and EGG features resulted in superior classification performance, especially when considering both genders simultaneously, highlighting the importance of feature fusion for effective voice pathology detection. For males, the accuracy surged to 92.75% with excellent performance across all metrics. Similarly, for females, the accuracy increased to 89.80% with notable improvements in other performance indicators. The classifier achieved remarkable accuracy of 96.08% when considering both genders, with excellent performance across all metrics, indicating the performance of the combined feature set in gender-agnostic classification.

The classifier performed well in detecting voice pathology disorders, showing high accuracy, specificity, recall, precision, and F1-score. These findings suggested that integrating multiple features can significantly enhance the accuracy and reliability of voice pathology detection. The classification model's consistent low MCR across all feature types underscored its robustness. Moreover, the utilization of GWOESCA-SVM demonstrated its efficacy in optimizing classifier performance across different feature sets and gender categories.

4.4 Performance comparison with standard SVM

To confirm the developed characteristics of GWOESCA, a typical SVM is used to compare the system's classification effectiveness. The results are shown in Table 5 without feature selection and parameter tuning.

4.4.1 Speech features

For males, classification accuracy was 79.22%, with moderate specificity, recall, precision, GM, F1-score and MCR. For females, the accuracy was slightly lower compared to males, with similar performance in specificity, recall, precision, GM, F1-score and MCR. Combined features demonstrated higher accuracy compared to individual genders, indicating a potential synergistic effect.

Table 5. Performance of the developed system without feature selection and parameter optimization

4.4.2 EGG features

For males, the system attained an accuracy of 71.18%, with specificity and recall values around 69.20% and 73.08%, respectively. Other metrics are moderate. For females, lower performance compared to males, with an accuracy of 66.75% and similar specificity and recall rates. Combined features demonstrated improved accuracy compared to individual genders, achieving 80.78%.

4.4.3 Combined features

Combined male features exhibited increase in accuracy compared to individual feature types, with 83.92%. However, the specificity is lower, while recall and precision were higher. Combined female feature gave a high accuracy of 81.96%, with balanced specificity and recall rates.  Combined male and female features demonstrated better results compared to individual genders, with an accuracy of 89.22%.

From the analysis, it can be noted that system without feature selection and parameter optimization showed moderate performance across all feature types and genders categories. However, there is room for improvement, particularly in optimizing parameters and selecting relevant features, which could potentially enhance the performance metrics.

Empirical findings demonstrate that the GWOESCA-SVM gave an excellent outcome exceeding that of standard SVM classifier, by registering classification accuracy of 88.82% for speech features, 84.71% for EGG features and 96.08% for combined features and for all cases of 10-cross fold validation with feature selection and parameter optimization. The efficacy of the introduced framework employing HNIA and ML has been demonstrated to yield superior results, affirming that GWOESCA-SVM adeptly discriminates between healthy and pathological voice samples. Pictorial representation of the developed SVM and GWOESCA-SVM is depicted in Figure 5, Figure 6 and Figure 7 for male, female and combined samples, respectively.

tu_pian_16.png

(a) Accuracy

tu_pian_17.png

(b) Specificity

tu_pian_18.png

(c) Recall

tu_pian_19.png

(d) Precision

tu_pian_20.png

(e) GM

tu_pian_21.png

(f) F1-Score

Figure 5. Performance comparison of the proposed system with standard SVM for male samples

tu_pian_22.png

(a) Accuracy

tu_pian_23.png

(b) Specificity

tu_pian_24.png

(c) Recall

tu_pian_25.png

(d) Precision

tu_pian_26.png

(e) GM

tu_pian_27.png

(f) F1-Score

Figure 6. Performance comparison of the proposed system with standard SVM for female samples

tu_pian_28.png

(a) Accuracy

tu_pian_29.png

(b) Specificity

tu_pian_30.png

(c) Recall

tu_pian_31.png

(d) Precision

tu_pian_32.png

(e) GM

tu_pian_33.png

(f) F1-score

Figure 7. Performance comparison of the proposed system with standard SVM for mixed samples

4.5 Performance comparison with other optimization algorithms

To verify the property of developed GWOESCA-SVM, the model’s performance was compared with that of the ESCA-SVM and GWO-SVM is shown in Figure 8. The GWOESCA-SVM method outperformed both GWO-SCA and ESCA-SVM in terms of all metrics, indicating its superiority in classification performance. While GWO-SVM closely followed GWOESCA-SVM, showing strong performance across all metrics, the ESCA-SVM demonstrated lower performance compared to both the GWOESCA-SVM and GWO-SVM across various metrics.

tu_pian_34.png

Figure 8. Performance compared with other optimization algorithms

To further validate the effectiveness of the proposed GWOESCA-SVM framework, we extended the experimental comparison by incorporating two additional nature-inspired optimization-based classifiers, namely, Particle Swarm Optimization (PSO)-SVM and Genetic Algorithm (GA)-SVM. These models were implemented using the same dataset and fusion strategy, with parameter tuning and feature selection optimized via PSO and GA, respectively. The experimental setup maintained consistent parameters for fair comparison: a population size of 30 and a maximum of 500 iterations. The classification accuracy was used as the fitness function evaluated using 10-fold cross-validation.

Table 6 presents the classification accuracy, F1-score, and MCR for all three models. As shown, GWOESCA-SVM outperformed both PSO-SVM and GA-SVM, achieving a maximum accuracy of 96.08% and the lowest MCR of 0.04. While PSO-SVM showed moderate performance, GA-SVM trailed slightly behind. This indicates that the hybrid exploration–exploitation behaviour embedded in GWOESCA is more effective in optimizing both the feature subset and SVM parameters for this complex, high-dimensional voice pathology dataset. These results are further visualized in Figure 9, which graphically compares the key performance indicators across all three optimization strategies.

Table 6. Performance comparison with other optimization algorithms

Figure 9 presents a comparative view of three optimization-based classifiers, namely, proposed GWOESCA-SVM, PSO-SVM, and GA-SVM, evaluated on core performance metrics. From the graph, depicted in Figure 9, it is evident that GWOESCA-SVM outperforms the other two techniques across all indicators. Specifically, it achieves the highest accuracy (96.08%), along with a near-perfect F1-score (96.15%), and maintains the lowest MCR (0.04). In contrast, both PSO-SVM and GA-SVM perform reasonably well, but with relatively lower precision and slightly higher error rates. This comparison confirms the superiority of the hybrid GWOESCA optimization approach, particularly in handling fused speech and EGG features, which are high-dimensional and heterogeneous. The result demonstrates that GWOESCA offers a more balanced and effective search strategy, leading to better model generalization. These findings strengthen the claim that the proposed method provides a more robust solution for voice pathology classification, suitable for real-world diagnostic applications.

tu_pian_35.png

Figure 9. Comparison of optimization-based classifiers

4.6 Comparison with the past approaches

In using HNIA with ML, the project seeks to develop an automated approach for detecting speech pathology. As reported in Table 4, the introduced framework, GWOESCA-SVM has attained a highest classification accuracy for all cases when compared to standard SVM.  Nevertheless, the performance is remarkably boosted when merged features are used, features are chosen by GWOESCA and parameters of the SVM is tuned by GWOESCA. The GWOESCA-SVM evidences its robustness and generalization capabilities considering all metrics in discriminating healthy from pathological voice samples.

Table 7 provides a comparative outcome of the simulation results for voice pathology detection between the introduced system and the former methods considering classification accuracy. To ensure a fair and genuine comparison of findings, only earlier approaches that utilized SVD data and focus on the /a/ vowel for investigation are considered. Voice pathology detection using GMM attained a classification rate of 67% [14]. El Emary et al. [15] categorized voice samples into healthy and pathological one using SVM, yielded an accuracy of 82.37%. Voice pathological detection system based on temporal derivatives and SVM classifier gave an accuracy of 86% [16]. Amara et al. [17] differentiated healthy from pathological voice samples using MFCC and SVM and obtained 95.5% accuracy. An accuracy of 90% was attained using NBN classifier [18]. Verde et al. [19] utilized four signal features and SVM classifier, attaining a classification accuracy of 85.77%. Mohammed et al. [31] adopted Resnet 34 for differentiating healthy from pathological samples, yielding a classification accuracy of 93.72%. Voice pathology classification approach using MLP achieved a classification rate of 82,77% [13]. Omeroglu et al. [20] investigated the power of standard SVM in categorizing voices samples into healthy and pathological ones and obtained 90.10% accuracy. Islam et al. [23] used CNN for voice categorization and yielded classification accuracy of 80.3%. Deep learning method-based voice detection methods yielded 73.83% accuracy [22]. Ksibi et al. [21] designed a hybrid approach by combining CNN-RNN for voice classification and reached an accuracy of 88.33%. The relevant features chosen by the GWOESCA tuned SVM resulted in a highest classification rate of 92.75% for male speech features, 89.80% for female features and 96.08% for combined male and female data.

The proposed method outperformed previous models by classification accuracy for both individual genders and the combined dataset. Several methods achieved moderate accuracy, as reported by Ahmed et al. [17], Mohammed et al. [31], and Hossain et al. [32]. Methods proposed by Martínez et al. [14] and Xie et al. [22] showed relatively lower accuracy compared to the others. The proposed GWOESCA-SVM method demonstrated significant advancement in accuracy compared to previous approaches, making it promising technique for voice pathology detection.

The current experimental design is limited to the Saarbruecken Voice Database (SVD), which was chosen due to its high-quality recordings, availability of both speech and Electroglottography (EGG) signals, and extensive labelling of multiple pathological classes across gender.

However, the cross-dataset validation plays a critical role in confirming the robustness and applicability of any machine learning system in real-world scenarios. Unfortunately, as of this study, publicly available databases with synchronized speech and EGG signal recordings are limited, and most alternatives (e.g., MEEI, AVPD) do not provide both modalities or are restricted in terms of pathological diversity or demographic distribution. That said, to partially address this limitation and are as follows:

• We performed extensive 10-fold cross-validation within the SVD dataset across three different conditions: male-only, female-only, and combined gender samples, using speech-only, EGG-only, and hybrid features. This layered evaluation framework was designed to simulate variability and ensure the model generalizes across gender and feature types.

• Additionally, we compared our model’s performance against several prior state-of-the-art methods (depicted in Table 7), many of which also relied on the SVD database. This fair benchmarking allows direct assessment of improvements brought by the GWOESCA-SVM framework.

Table 7. Performance comparative study of GWOESCA-SVM with former methods

4.7 Performance variations between genders

This subsection analyses the difference in classification outcomes between male and female voice samples, based on both numeric metrics and confusion matrices.

tu_pian_36.png

Figure 10. Confusion matrix for male samples

tu_pian_37.png

Figure 11. Confusion matrix for female sample

To further investigate the performance variations between genders, we analysed the confusion matrices for male and female samples, as shown in Figures 10 and 11, respectively. In the male dataset (Figure 10), the system correctly identified 117 pathological and 115 healthy samples, with only 18 misclassifications. This reflects the relatively higher uniformity and clarity of acoustic patterns in male voices, making pathological deviations more detectable. In contrast, the confusion matrix for female samples (Figure 11) reveals a slightly lower classification accuracy, with 30 healthy samples misclassified as pathological and 24 false negatives. This increase in misclassification can be attributed to the greater variability in pitch, voice modulation, and EGG signals in female subjects, which tends to blur the decision boundaries between healthy and pathological classes. These visual insights reinforce the numeric performance metrics and highlight the importance of considering gender-specific signal characteristics when designing robust voice pathology detection systems.

4.8 Computational efficiency and practical feasibility

To address the computational efficiency and real-world applicability, quantitative measurements of training and inference times for the proposed GWOESCA-SVM model is determined. These measurements were conducted on the same hardware configuration used for all experiments.

(a) Training Time: For the combined feature set (speech + EGG), the average training time per 10-fold cross-validation cycle was approximately 135 seconds. This includes time for feature selection and SVM parameter optimization using GWOESCA (population = 30, iterations = 500).

(b) Inference Time: Once trained, the average classification time per test sample was approximately 0.012 seconds, indicating suitability for near-real-time deployment in diagnostic tools.

(c) Discussion: While the hybrid nature-inspired optimization phase introduces some computational overhead during training, this cost is acceptable given the one-time offline training requirement. The inference phase remains lightweight, making the system practical for real-time or embedded voice assessment scenarios in clinical or telemedicine environments.

4.9 Robustness to noise and clinical deployment considerations

(a) Robustness to Noisy or Low-Quality Signals: Although the SVD used in the proposed study comprises clean, high-quality recordings, real-world voice data, especially from clinical or remote environments, is often subject to background noise, recording artifacts, or variable microphone quality. While our current system does not explicitly include noise-handling mechanisms, it maintains a degree of robustness due to the following:

• The use of multiple feature types (e.g., MFCC, HNR, ZCR, spectral descriptors) provides redundancy and resilience to mild signal distortions.
• The GWOESCA optimization process inherently selects robust features that remain consistent across samples, which can reduce the impact of noise.
• EGG signals, being physiological rather than acoustic, are less susceptible to ambient noise, offering a stable complementary signal source.

(b) Clinical Deployment Feasibility: The proposed GWOESCA-SVM framework was designed with low inference cost and lightweight deployment potential in mind. Once trained, the classifier operates with an average inference time of 0.012 seconds per sample, as discussed in Section 4.8. This makes it well-suited for real-time screening applications, including, (i) Outpatient voice clinics, (ii) Telemedicine consultations, and (iii) Smartphone-based remote voice analysis tools.