VocalCare: Convolutional Neural Network Model for Laryngeal Disorder Detection

One of the biggest challenges facing contemporary healthcare systems is the early and precise detection of several diseases. Traditional binary classification algorithms are inadequate in handling real-world clinical circumstances when patients may present symptoms matching to numerous illness categories. In this paper, an automated multiclass classification framework for machine learning-based disease detection and differentiation is presented. The comprehensive analysis of major features was used to catch the differentiating characteristics for different forms of clinic and clinical data. In forecasting various types of diseases, 4 different machine learning models were tested such as support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF) and convolution neural networks (CNNs). The suggested model is proved by experimentations of large numbers of samples that demonstrated high-quality predictions and reliability and indicates the ability for assisting various clients with accurate and fast detection of diseases.

The human larynx is a functional structure with a complex shape located in the neck and plays vital roles in respiration, sound, and airway protection. Potential states of the human glare can greatly impact one's ability to breathe, speak, and sound fully. For example, laryngeal papillomatosis, vocal fold nodules, vocal fold polyps, vocal fold paralysis and/or cancer are potential outcomes of the human glare, and therefore, can have a negative social or psychological effect on the individual as well as the community. Therefore, the timely and accurate detection of laryngeal disorders is paramount to achieving appropriate and correct medical treatment; subsequently improving and enhancing the likelihood of positive medical treatment outcomes and minimizing the long-term effects of laryngeal dysfunction. Laryngeal disorders can be diagnosed using traditional diagnostic methods such as endoscopic, video stroboscopic, and laryngoscope procedures. These traditional diagnostic procedures have been successful however, they are expensive, time consuming, very confidential, and largely subjective based on a medical practitioner's skill set. Additionally, due to restricted access to specialized laryngology services, diagnosis and treatment may be delayed in distant or resource-constrained places. These problems show how important it is to have automated, non-invasive, and affordable ways to find laryngeal disorders early.

Deep learning has made major strides in artificial intelligence, which has greatly increased the ability to analyze acoustic signals. This has altered the field of medical diagnosis. Convolutional neural networks CNNs, which are a sub-class of deep neural network (DN) algorithms, have demonstrated high levels of success in the classification of images and audio signals through their ability to learn structural features from unprocessed input data with regard to the diagnosis of laryngeal diseases, it is common to convert the sound waves collected from a patient’s speech, into visual representation such as a spectrogram, or some other representation in time-frequency space. Within these images are subtle and difficult-to-identify patterns that can be effectively learned and found using the convolution neural network (CNN)-based analysis model. Numerous previous studies have highlighted the significant potential of using AI to support clinical decision making, specifically with regard to the use of machine learning and deep learning to detect speech disorders. In general, these studies have focused on extracting the features that will be used to classify a patient’s speech (e.g. pitch, jitter, shimmer, Mel-Frequency Cepstral Coefficients (MFCCs), followed by using traditional classifiers, such as a RF or SVM, to classify the extracted features of a patient’s speech.

While there are certainly promising results from the use of these techniques, they are primarily based upon the use of manually defined features and because of this, they may have difficulty generalizing across different patient populations. CNNs on the contrary, provide the ability to automatically extract the required features from the input (i.e. patient’s speech) and provide a greater degree of robustness and predictability, thereby improving the accuracy of the predictions.

In this paper we introduce a novel Speech Recognition System based on CNNs for identifying Laryngeal Disease through recorded Speech Data. This method will use spectrogram "visualizations", which are a representation of sound, in conjunction with CNNs to extract the distinguishing features of Speech Data and separate the two types (diseased vs. healthy). This system was designed to be: scalable, non-invasive, allow physicians to diagnose their patient's condition sooner and reduce the need for invasive diagnostic measures, and to expand the reach of high quality medical care.The study discusses the potential benefits of incorporating AI technology into such fields as speech pathology and otolaryngology. This framework will help doctors to make clinical decisions and will provide novel telemedicine applications and remote monitoring for speech health, offering a reliable and automated diagnostic tool for the diagnosis of laryngeal illnesses. The comprehensive testing and analysis of the CNN-based model has demonstrated the usefulness of the model in diagnosing laryngeal illness and can potentially transform present diagnostic techniques and improve patient outcomes. This complete chain makes it possible for CNN to perform reliable automatic acoustic assessment of first vocal disorders.

This paper further highlights the possible benefit of the application of AI in speech pathology and otolaryngology. The proposed architecture supports clinical decision making, and provides a gateway for telemedicine applications, remote monitoring and continuous speech health assessment by providing a reliable automated tool for diagnosis of laryngeal diseases. We show that the CNN-based model can be effectively used for the diagnosis of laryngeal disorders by applying the model to a large number of patients and analyzing the results. It has the potential to change the current diagnostic procedures and enhance patient outcomes. The whole cycle allows the CNN to perform reliable, automated acoustic analysis for preliminary vocal pathology screening. Speech pathology identification has been extensively studied using standard signal processing techniques as well as recent deep learning approaches. Recent approaches have used CNNs to learn discriminative features from the speech input, whereas the former approaches focused mostly on acoustic and spectral properties.

The novelty of the proposed Vocal Care system is the construction of an intelligent CNN based framework for autonomous laryngeal health monitoring using human speech data. The proposed strategy is non-invasive, cost-effective and detects speech abnormalities at an early stage, in contrast to the conventional clinical diagnostic methods based on invasive techniques such as laryngoscopy and manual analysis by experts. The proposed method adopts the advanced machine learning and deep learning methods. The Mel-Frequency Cepstral Coefficient (MFCC) is used to automatically extract the discriminative acoustic features from the speech signals, which minimizes the dependence on the manual feature extraction methods and clinical interpretation. The CNN-based framework learns important speech patterns related to laryngeal anomalies in an efficient manner, resulting in higher accuracy and reliability in the early diagnosis of vocal disorders. A comparison analysis of standard ML methods with the proposed CNN model reveals that the deep learning based framework offers better classification accuracy, robustness and faster detection performance for identification of laryngeal anomalies.

Furthermore, the Vocal Care system can also help improve access to healthcare services by allowing patients to remotely and continuously monitor their speech health especially in rural and underserved areas. Early stage diagnosis is also possible with the suggested approach thereby reducing the chance of severe speech problems by appropriate intervention.

3.1 Dataset details

The experiments in this study were conducted using the Saarbrücken Speech Database (SVD), a publicly available pathological speech dataset widely used for speech disorder analysis research. The database contains recordings from healthy individuals as well as patients diagnosed with different pathological speech conditions.

A total of 1,037 speech recordings were utilized in this work. The dataset includes sustained vowel phonations and pathological speech samples collected under controlled recording conditions. Our study used speech recordings obtained from the SVD, which contains sustained vowel sounds (/a/, /i/, /u/) from both healthy people and patients with laryngeal pathologies, such as paralysis and nodules. Each recording is captured at 50 kHz and has a 16-bit resolution. The vowels are /a/, /i/, and /u/ are recorded in this set together with an appropriate sentence. When evaluating a patient's speech quality, vowels are preferred above linguistic artifacts. To conduct the experiments on in our experimental testing, Specifically, 477 pathological speech (392 samples of Vox Senilis & 85 samples of Laryngocele) and 560 healthy samples have been selected. The dataset was divided into training and testing subsets using stratified sampling to preserve class distribution across both partitions, with 80% used for training and 20% for testing. Experiments were conducted to evaluate the performance of the proposed model in accurately classifying multiple diseases across different categories.

3.2 Feature extraction

MFCC: This is popular feature techniques for extraction in the field of speech recognition and audio processing. MFCCs are especially popular in tasks like as speech recognition, speak identification, and music data retrieval. Following are the steps for Mel-Frequency Cepstral Coefficients.

Pre-processing: The speech signal first undergoes pre-processing to eliminate areas of noise and silence. The signal is often broken into brief frames of 20 to 40 millisecond duration, with only a slight overlap among consecutive frames.

Mel-scale Filter Library: The FFT is used to determine the power spectrum of every frame. The power spectrum is then routed through a bank of filter spaced out on the Mel-scale to better simulate the non-linear perception of frequencies by the human auditory system. The Mel-scale is a pitch perception scale based on the human ear's reaction to various frequencies.

Logarithm: The filter bank energies' logarithm is computed. This stage compresses the spectrum's dynamic range and simulates the non-linear human sense of loudness.

Discrete cosine transform (DCT): The energies of the log filter bank are subjected to a discrete cosine transformation. DCT de-correlates motion coefficients.

Coefficient Selection: After applying the DCT, typically a selection of the resulting coefficients is made, leaving out the higher-frequency coefficients that hold less pertinent information. The selected coefficients then serve as the final MFCC characteristics.

3.3 Multiclass classification models

In this research, we used MFCCs to classify the speech waveform data into three categories; Laryngeal pathologies, Normal speech, and Vox senilis speech. MFCCs are frequently used to analyze speech disorders and efficiently model the perceptual properties of speech sound produced by the human body. Consequently, we developed classification algorithms based on the MFCC characteristics identified to classify the speech waveform data into the three categories: SVM, KNN, RF and CNN were developed, and we analysed their performance in differentiating normal and sick speech through performance measures such as accuracy, sensitivity and specificity, as well as the F1 score, through confusion matrix-based performance analyses. The laryngeal sample and the vox senilis (senior citizens) sample, had very slight misclassifications between both samples; however, there were a large degree of separation between both samples Laryngeal & Normal than vice versa (Normal & Vox Senilis) & thus, they had different classification performances; therefore, the misclassifications could be attributable to the sound spectra and auditory characteristics that were similar; but still the overall results indicate that MFCC acoustic measurements are a valid representation of the speech produced by people with Laryngeal disorders such as those produced in a normal vocal cord or Vox Senilis such as with senior citizens and have been helpful in diagnosing & assessing speech disorders.

Three techniques for machine learning were used:

•SVM: Models non-linear class boundaries using an RBF kernel.

•KNN: This method uses Euclidean feature space similarity to classify samples.

•RF: Reduces overfitting and increases prediction stability by combining decision trees.

CNN: A neural network that uses an MFCC representation for speech samples and is taught via deep learning. This deep learning model is able to use these MFCC representations of sound to learn multiple levels of spectro-temporal features associated with the laryngeal speech, such as normal speech, abnormal speech and senile speech. Ramitha et al. [18] used a CNN model to accurately identify very small sounds associated with the laryngeal speech.

3.4 Evaluation metrics

Performance was measured using various metrics, including accuracy, precision/recall, specificity, and F1-score, while k-fold cross-validation was used to improve generalization. The robustness of the proposed classification method for speech pathology was tested by evaluating the performance of each model using a strategy of 5-fold cross validation. The data set of available speech was divided at random into five semi-equal size subsets; the first 4 subsets were used for training and the remaining subset was kept out for testing (guaranteeing that each speech sample was examined precisely once). By combining all of the results from each fold, the overall performance of the classifier could be assessed without bias on new speech samples.

The following evaluation metrics measured the diagnostic capabilities of the model:

Accuracy: Accuracy is an estimate of how accurate the classifier is based on how a lot of a speech samples had been accurately classified as being diseased or not diseased. Even though accuracy provides a general view of a classifier's performance, it can be significantly affected by the class imbalance (i.e., fewer samples in the speech disorder classes) commonly seen in most speech pathology data sets.

Sensitivity (Recall) indicates how well the model distinguishes between various speech conditions (example: Disease and No Disease) by dividing the number of True Positives for each speech condition. High sensitivity is especially important when evaluating the speech for clinical reasons since high sensitivity indicates that the speech evaluation system has a high probability of identifying pathological conditions therefore reducing the number of missed diagnoses.

Specificity measures how well the classifier distinguishes between Non-pathological speech samples by dividing the number of True Negatives (non-pathological speech samples) by the total number of Non-pathological samples. This is important for lowering the number of false alarms and making sure that normal speech samples aren't wrongly classified as pathological.

F1-Score: The F1-score is a balanced performance evaluation that is especially useful for imbalanced speech datasets, where some vocal disorders may be underrepresented. It is represented by the harmonic mean of precision and recall. The F1-score gives an all-encompassing evaluation by taking false positives and false negatives into account at the same time.

Collectively, these measures give a detailed and clinically meaningful analysis of the efficacy of the proposed approach in categorizing speech pathologies.

3.5 Speech based laryngeal abnormality detection system

Figure 1 shows that this method uses a CNN to analyze acoustic data taken from recorded speech in order to classify speech signals into Normal or Abnormal (Laryngeal Disorder) categories.

1.png

Figure 1. Speech based laryngeal problem identification utilizing convolutional neural network (CNN)

3.5.1 Pre-processing and signal acquisition

A task that may be performed during the experiment is for the participant to extend their vocalization of one of the vowel sounds (a, i, or u) in order to provide a speech sample in a controlled environment. Normalization of each recording takes place, where the amplitude is corrected for any changes in both the volume of the speaker and the distance of the microphone from the speaker, so that the recordings are normalized. Prior to performing the frequency analysis on the continuous one-dimensional (1D) signal, frame blocking is used to separate the continuous 1D signal into a number of segments that overlap one another. An appropriate window function is then applied to each of the segments in order to reduce distortion.

3.5.2 Extracting features with Mel-Frequency Cepstral Coefficients

In order to create the needed two dimensional (2D) input required by the CNN, the MFCCs were used to create a 2D representation of the human ear's hearing perception. Each frame's frequency spectrum was calculated and passed through a Mel-frequency filter bank comprise of filters designed to highlight low frequencies and then the outputs were truncated as perceivable loudnesses. In order to obtain MFCCs, the DCT technique was used on the outputs from the previous stage. The resulting sequence of MFCC vectors thus forms a 2D feature matrix (time × coefficients), which is required for the CNN's inputs.

The CNN extracts and learns discriminative speech patterns through two main stages:

3.5.3 Feature learning

The layered structure of convolutional networks allows for a reduced number of calculations and an improvement in the ability to generalize from images at the same time as allowing for the recognition of localized spectral patterns from filtered input. Convolutional filters identify localized spectral patterns in the input through a set of convolutional (convoluted) networks using Rectified Linear Unit (ReLU) activations, which add non-linearities to the output.

3.5.4 Classification

After being flattened into a 1D vector, the final pooled features are sent to fully connected (FC) layers, which combine the learned features to get a final conclusion. Each class's likelihood scores are produced by the output layer.

3.5.5 Classification results

Based on machine learning and deep learning classification algorithms with probabilities, the system offers a binary categorization:

3.5.6 Normal Speech (Healthy)

The normal class has a higher probability.

3.5.7 Abnormal Speech (Laryngeal Disorder)

Potential vocal pathology is indicated by a higher likelihood assigned to the abnormal class.

This CNN-based system efficiently transforms unorganized speech signals into structured representations. By employing automated acoustic pattern analysis, it enables robust, data-driven diagnosis of laryngeal diseases.

5.1 Normal, Vox Sensils, and Laryngocele results using Support Vector Machine, k-Nearest Neighbors, and Random Forest

This study shows that by experimenting with different combinations of vocal fold conditions (Normal, Vox Sensilis, Laryngocele) and using several different supervised learning algorithms (SVM, KNN, RF), it is possible to effectively classify pathological speech disorders. A dataset consisting of 560 healthy, 85 laryngocele, and 395 vox senilis was used in this study. The classifiers were tested using 5-fold cross-validation so that each speech sample was in the test set exactly once. A combination of MFCCs was created to use as input to the classifiers; jitter, shimmer, and HNR features were added to improve the discriminative capabilities of the classifiers. The classifiers will be trained to classify speech as healthy, laryngocele, or vox senilis types. The SVM classifier was tested using 5-fold cross-validation, with each fold used once as the test set and the remaining folds for training. This method allowed every sample in the collected data to contribute to the testing process.

Figure 3 indicates that confusion matrix obtained using the SVM classifier shows an overall accuracy of the class-wise accuracy of the SVM classifier after 5-fold cross-validation shows varying performance across the three classes. The Normal class achieved the highest accuracy of 86.4%, indicating that the model is highly effective in correctly identifying normal speech samples. The Vox senilis class was highly accurate, achieving an accuracy rate of 80.3%. There was, however, some misclassification of Vox senilis with other classes due to confusion regarding their characteristics; nonetheless, the Vox senilis class still performed relatively well overall. The Laryngozele class, on the other hand, was not as well detected, obtaining an accuracy of 58.8%. This suggests more difficulty in identifying Laryngozele conditions because it was frequently misclassified as Vox senilis.

Figure 3. Classification results obtained by Support Vector Machine (SVM) algorithm for Normal, Vox Sensilis, and Laryngocele class

Overall, the classifier achieved an excellent overall accuracy of 81.8%, but there was a lot of variation in class-level accuracy. The results suggest that reporting class-level accuracy in addition to total accuracy is important when evaluating clinical speech pathology classification methods, especially when classifying minority classes, which are generally more difficult to detect.

Figure 4 indicates the RF classifier is shown to provide excellent performance on the speech dataset and produces a strong overall result of 83.75% when tested with 5-fold cross-validation on the speech pathology dataset. This model successfully identifies "healthy" speech with a very high sensitivity of 92.68%. The model also is able to successfully recognize cases of "vox senilis" (elderly, aged) at a rate of 81.27%. It is important to note that there was confusion between the two classes and many of the classified cases were misclassified either as healthy or as vox senilis.

Figure 4. Classification results obtained by random forest (RF) classifier for Normal, Vox sensils and Laryngocele class

The results suggest that the confusion between the two classes may be attributable to a class imbalance and the fact that there is a significant overlap in the acoustic characteristics of the two classes. As illustrated in Figure 3, the RF model exhibit a significant increase in robustness and ability to discriminate between the two classes when compared to the simpler models especially with regard to the healthy and vox senilis. Figure 5 indicates KNN Classifier Solutions for Speech Recording Classification Shows ~80% Accurate Classification of Speech Recordings Overall. The KNN Classifier Shows High Performance in Classifying Healthy Speech, with a High Percentage of Healthy Speech at 86%. The KNN Classifier Shows Moderate Performance in Classifying Vox Senilis at 76.71%, Demonstrating the Overlap between the Normal Speech and Age-Related Speech. At the Lower End of the Classification Percentage (%of 48.91%), Laryngocele Is Not as Easily Classified by KNN than by Other Classes. The Figure 4 Supports the See that KNN Is a Good Example of Classification of Healthy Speech. Results of the Experiments Support that the Feature Set Described Above Will Enable the Modeling of the Characteristics of Pathologic Speech Using the Classifiers That Were Used for This Study. The Classification of Healthy Speech is Very High (and Accurate) Because of It Unique Acoustic Signature and the High Number of Included Healthy Speech in the Data Set and Vox Senilis However, Pathological Conditions Such as Laryngocele Were Not Easily Resolved Because of Limited Samples of Pathological Speech. and overlapping acoustic features with vox senilisThe study shows that using machine-learning based techniques in pathological assessment of speech has potential for being used in practice and highlights the need for appropriate feature selection and case mix in relation to diagnostic accuracy and generalization. It was determined that when using 40 MFCC features, Jitter, Shimmer and Harmonic to Noise Ratio (HNR) the accuracy of the three models improved considerably. The RF classifier had the highest accuracy of 83.17%, which shows its advantage for processing non-linear patterns.

Figure 5. Classification results obtained by k-nearest neighbors (KNN) algorithm for Laryngocele, Healthy and Vox sensils class

The next highest accuracy was obtained by the KNN method at 78.85% and is considered reliable in identifying speech that have similar acoustic properties. The SVM classifier also demonstrated a moderate accuracy of 76.92% and indicates some difficulty with classifying overlapping feature distributions. Overall, incorporating MFCC, Jitter, Shimmer, and HNR features proved highly effective for representing vocal characteristics, and the RF model demonstrated the most stable and accurate performance for classifying Normal, Vox Senilis, and Laryngocele speech samples.

5.2 Normal, results based on 5-fold cross-validation

This experiment employed five-fold cross-validation to see how well the machine learning models worked and how well they could be used in other situations. The dataset was split into five equal parts, and four of them were used for training and the fifth for testing in each round. Each fold was used as a testing set once over the five iterations of this approach. The resulting calculated performance indicators were also averaged over all five folds to give a precise and consistent evaluation of the models. Stratified splitting of the data was utilized to help preserve the original class distribution across each folds giving an accurate and reliable evaluation of the models’ performance. As shown in Figure 6, three different classification methods—SVM, RF, and KNN (k-nearest neighbor)—were used to assess fold-wise accuracy using 5-fold cross-validation. SVM's fold-wise accuracies were 84.62%, 83.65%, 82.69%, 79.33%, and 78.85%, with an average accuracy of 81.83% for all five folds. The fold-wise accuracies for RF were 83.65%, 84.62%, 81.25%, 84.62%, and 84.62%, with an average of 83.75% for all five folds. KNN (k-nearest neighbor) displayed variety in its fold-wise accuracies, reporting: 80.29%, 76.92%, 84.13%, 76.92%, and 79.33%, which gave KNN an average accuracy of 79.52%. While SVM and KNN offered modest accuracy and variability across the folds, RF had the best average accuracy and the most consistent results overall.

Figure 6. Calculated accuracy results 5 folds cross validation

The high average accuracy of the RF is due to its ensemble learning approach, which decreases overfitting as well as provides nonlinear representation of the features. Similarly, the variability of the SVM suggests that there is sensitivity with the data distribution and/or kernel selection, however, KNN’s accurate results illustrate that local neighborhood learning is consistent. Collectively, these results demonstrate that RF produces more accurate outputs than single classifiers method composed of SVM/RF/KNN classifiers when used to classify speech disorders. The performance of the SVM classifier across the five folds:

83.65% (fold 1), 84.62% (fold 2), 81.25% (fold 3), 84.62% (fold 4), and 84.62% (fold 5), for an average accuracy of 81.83%. The RF produced more consistently across folds with the following average fold accuracies of:

83.65% (fold 1), 84.62% (fold 2), 81.25% (fold 3), 84.62% (fold 4), and 84.62% (fold 5), resulting in an overall average accuracy of 83.75%. Finally, while the KNN exhibited a much greater variance in accuracy values. The five folds achieved accuracies of 80.29%, 76.92%, 84.13%, 76.92%, and 79.33%, which added up to an overall average accuracy of 79.52%. In general, RF performed the best. SVM and KNN both did okay, but there were notable differences between the folds.

The RF works better than other classifiers due to their ensemble learning (which minimizes the risk of overfitting by creating multiple trees) and its ability to find nonlinear relationships among features; however, the results from SVM show a great deal of variability based on how the algorithm reacts to the data set and how the specific kernel was selected. KNN produces results that are more stable and, therefore, indicate that KNN successfully learns local neighborhoods The results from our study indicate that RF provides more stable and reliable predictions related to speech issues than SVM or KNN when using single classifiers compared to having ensembles of classifiers working together.

5.3 Deep learning approach using Convolutional Neural Network

The study involves trials utilizing a CNN to evaluate its capacity to categorize vocal disorders based on audio characteristics. The dataset was made up of .wav files organized into three categories: Normal, Vox Senilis, and Laryngocele. Key speech features, including MFCCs, Jitter, Shimmer, and HNR (Harmonics-to-Noise Ratio), were extracted from these audio files. The CNN model was trained using an 80-20 train-test split, where 80% of the data was used for training and 20% for testing. 1D CNN model was trained for 50 epochs on the extracted features. Below is a summary of its training and validation accuracy.The training accuracy exceeded 98%, and validation accuracy stabilized around 88–91% from Epoch 30 onward and achieved best Validation Accuracy is 90.16%.

Figure 7 shows the training accuracy exceeded 98%, and validation accuracy stabilized around 88–91% from Epoch 30 onward and achieved best Validation Accuracy is 91.19%.

Figure 7. Convolution neural network (CNN) model training and validation accuracy

Epochs 1–10: Steady improvement in accuracy (from 68% to 84%)

Epochs 10–30: Model stabilizes (87 to 89% validation accuracy)

Epochs 30–50: validation accuracy peaks 90.16% shown in Among all models and shown in Figure 8 in confusion matrix, CNN achieved the highest accuracy of 90.16%, followed by RF with 83.17%, making CNN the most effective model for this classification task. The CNN model good performs all traditional machine learning models, demonstrating its ability to effectively capture complex temporal patterns in speech features.

Figure 8. Classification results obtained by using convolution neural network (CNN) for Normal, Vox sensils and Laryngocele class

The results in Table 1 show that while ensemble methods, like RF, give high accuracy for majority classes, the CNN offers the best balanced and strong diagnostic performance across all categories. With an accuracy of 90.16% and an F1-score of 85.80%, the CNN stands out as the most effective tool for automated laryngeal pathology detection in this study.

Table 1. Class-wise performance analysis using convolution neural network (CNN)

Successful multiclass classification outcomes were attained from the proposed model for the three classification types of laryngocele, normal, and vox senilis voice samples. The highest recall value (95.71%) was associated with the normal classification in terms of the three classes demonstrating exceptional detection capability. The highest precision value (94%) was associated with the vox senilis class demonstrating that the model generates highly accurate classifications. The combined precision values, recall values, and F1-score values provide additional evidence of the classification systems' robustness.

The proposed CNN framework achieved an overall classification accuracy of 90.16%.

5.4 Comparative performance analysis of different model using features Mel-Frequency Cepstral Coefficient, Jitter, Shimmer and Harmonic to Noise Ratio

The experimental results show in Figure 9 demonstrate that the use of additional acoustic features alongside MFCC improved the overall classification performance of the proposed system. Using MFCC features alone, the SVM, RF, KNN, and CNN classifiers achieved accuracies of 76.25%, 77.00%, 78.00%, and 89.00%, respectively. After incorporating Jitter, Shimmer, and HNR features with MFCC, the classification accuracies increased to 81.83% for SVM, 79.52% for RF, 83.75% for KNN, and 90.16% for CNN. The improvement in performance indicates that the additional acoustic features provide complementary information related to vocal instability and noise characteristics, which enhances the model’s ability to distinguish pathological speech conditions more effectively.

Figure 9. Comparative analysis using Mel-frequency cepstral coefficient (MFCC) and MFCC+Jitter+shimmer +hnr

Table 2 shows that the results indicate RF consistently performed better than both SVM and KNN in overall classification accuracy and specificity. This demonstrates its effectiveness for multi-class pathological speech detection. The high specificity of 94% in the CNN model highlights its strength in correctly identifying healthy samples compared to diseased ones, such as laryngocele and vox sensilis.

Table 2. Comparative performance analysis of different models using Mel-frequency cepstral coefficient (MFCC), Jitter, Shimmer, hnr

5.5 Clinical relevance

According to the experimental evaluation, the suggested CNN-based Vocal Care framework performed reliably in differentiating between healthy and pathological speech samples, achieving a classification accuracy of 90.16%, sensitivity of 85.53%, specificity of 94%, and F1-score of 85%.From a clinical standpoint, the suggested Vocal Care framework offers an automated, non-invasive method for early laryngeal disease diagnosis utilizing speech data. The system's capacity to successfully detect aberrant speech situations while reducing misclassification is demonstrated by the excellent specificity and sensitivity values.Furthermore, the proposed CNN-based system can support healthcare professionals by enabling rapid preliminary screening in remote and resource-limited healthcare environments. The lightweight architecture and automated analysis capability also make the framework suitable for integration into telemedicine and mobile healthcare applications.