A review on voice pathology: Taxonomy, diagnosis, medical procedures and detection techniques, open challenges, limitations, and recommendations for future directions

1 Introduction

Normal voice is the result of the interaction of the larynx with the pulmonary air pulses, which sets the actual movement of the vocal fold toward the midline and results in the production of sounds, either aperiodic or periodic. Voice disorder is defined as any sort of abnormality that deviates from the characteristics such as “loudness, quality, pitch and/or vocal flexibility,” compared with voices from the same age group, sex, and social group [1,2]. Twenty percent of young adults and children aged 3–21 years had voice problems based on a newly published survey from the National Center for Education Statistics. The American Speech-Language-Hearing Association found that voice impairment is caused by a defect in the human voice generation system [3].

Invasive surgical treatments are frequently performed to diagnose vocal abnormalities. Doctors put a probe into the mouth during endoscopic procedures, such as laryngoscopy, laryngeal electromyography, and stroboscopy. These operations are painful and can traumatize patients. Finding alternatives to these surgical treatments has been the subject of extensive research. One of these alternatives is utilizing voice signal processing to detect vocal pathology [4]. Extraction and analysis of voice features are required for vocal pathology detection using voice signals. Voice samples were taken in a specific environment. The extracted voice features from the voice signals were then examined. Voice samples were then divided into two groups: normal and abnormal. The right technology must be chosen carefully considering voice signal identification. The classification algorithm is the most widely used technique to identify and differentiate between two or more subjects [5]. Machine learning has shown efficient performance in various medical domains, including cancer diagnosis and classification [6], 3D brain reconstruction [7,8], and Alzheimer’s disease [9,10,11].

The previous research has shown that the accuracy level of a classifier is significantly dependent on this measurement. However, numerous concerns and challenges are associated with pathology detection approaches based on voice signals. Several issues such as (1) selection of adequate voice features and (2) identification of an appropriate classifier [5] are also important. The current study aims to present and discuss an exhaustive analysis of various voice abnormalities, machine learning, voice disorder databases, classifiers, and feature extraction methods. Moreover, this study has been employed for the development of a mechanism that works automatically for the classification and detection of voice pathologies. The primary contributions of this work are presented as follows.

• This review covered voice pathology taxonomy, diagnosis, medical procedures and detection techniques, open challenges, limitations, and recommendations for future directions to provide clear background for patients.

• This study presents medical conditions for voice pathology with their categories, and this includes swallowing difficulties, speech defects, language defects, cognitive communication defects, and social communication defects.

• Standard databases were used in most articles that we have reviewed in this article, including Massachusetts Eye and Ear Infirmary (MEEI), Saarbruecken Voice Database (SVD), and Arabic Voice Pathology Database (AVPD), and also, highlighted classes, features, and main purpose for voice pathology identification.

• This study focuses on voice pathology feature extraction, especially the speech analysis, extended feature vector made of static and dynamic features, and converts these extended feature vectors into more solid vectors before passing them on to the recognizer.

• The current study demonstrates the importance of AI approaches, such as machine and deep learning, which are used to classify the cases into healthy or pathological.

• This study also demonstrates that Internet of things (IoT) is frequently only partially and correctly utilized in the healthcare domain due to a significant gap between the data used and the ability to process and analyze the approach. The largest challenge for each patient, particularly those living in rural places, lies in their inability to get doctors and remedies in emergencies. Therefore, an innovative system based on the IoT is required in these areas.

The following is a breakdown of the structure of this article. Section 2 demonstrates an example of voice pathologies. Section 3 explains the voice pathology database. Section 4 discusses approaches for extracting voice pathology features. Section 5 presents voice pathology classification algorithms. Section 6 explains IOT voice pathology. Section 8 concludes this article.

2 Voice disorders

This section summarizes the voice disorders as follows:

3 Voice pathology database

The voice pathology database is a critical component of the automatic voice disorder detection (AVDD) system. Samples of healthy and disordered voices are included in the dataset. These samples may include sustained vowel phonation or continuous speech. Several common databases such as the MEEI, the SVD, and the AVPD are used in most studies [14].

3.1 Massachusetts eye and ear infirmary

A database was created by the MEEI lab for voice and speech. It contains around 1,400 voiced, sustained vowel/a/samples, as well as the initial half of the Rainbow Passage. Kay Elemetrics is the company that sells this database [15,16]; it was captured in two different settings. Normal samples were experimented with at 50 kHz, while pathological samples were tested at 25 or 50 kHz. A spreadsheet with clinical and personal information from the individuals as well as the findings of the acoustic analysis of the recordings collected from Kay’s MultiDimensional Voice Program (MDVP) are also included in the database. The Computerized Speech Lab (CSL) of Kay was also used to make the recordings under similar acoustic settings. Each participant was instructed to generate a sustained phonation of the vowel/ah/ at a comfortable pitch and loudness for at least 3 s. The technique was performed three times for each participant, and the best sample was chosen for the database by a speech pathologist [17]. Despite its popularity, this database has several drawbacks. The database is used in most investigations of vocal pathology identification and classification and utilizes different settings and frequencies of voice samples to capture healthy and disordered voices (Kay Elemetrics confirmed this data). A variety of vocal disorders could also be detected by analyzing changes in the voice muscles that can activate and increase voice efficiency [15] (Figure 1).

Figure 1

Three types of signal examples from the MEEI database [15].

3.2 Saarbruecken voice database

SVD is an available free database [18] maintained by the Institute of Phonetics at Saarland University. This database includes/a/,/i/, and/u/sustained vowels with several intonations, including low-high-low, low, normal, and high, as well as a spoken sentence in German, “Guten Morgen, wie geht es Ihnen?,” which translates to “Good morning, how are you?” These characteristics contribute to the remarkability of this database. The recorded voices of the SVD database were all sampled at 50 kHz with a 16-bit resolution. This database is also new, and only a small number of studies have used it in the detection of voice disorders [15]. Files can be downloaded from the website provided in [18]. The voice and EGG signals are saved for each recording. Both signals may be exported in their native file formats (NSP and EGG) as well as WAV, thus creating a zip file on the server containing the chosen files. The procedure might take several minutes. Finally, a link to download the produced zip file appears [18] (Figure 2).

Figure 2

Voice and EGG signal examples from the SVD database [13].

3.3 Arabic voice pathology database

Samples of this database were recorded at the Communication and Swallowing Disorders Unit (King Abdul Aziz University Hospital, Riyadh, Saudi Arabia) during several sessions in a sound-treated environment by skilled phoneticians using a defined recording methodology [15,19]. The gathering of the database was one of the key responsibilities of a 2-year initiative supported by the National Plans for Science and Technology (NPST) in Saudi Arabia. Database protocol was created in such a way that would avoid the flaws of the MEEI database [15,20]. The database contains various samples, including speeches from individuals with vocal-fold diseases, sustained vowels, as well as recordings from those with normal speech. Healthy and ill vocal folds were recognized after a clinical examination with laryngeal stroboscopy. On a scale of 1–3 severity, the frequency of voice problems was graded in cases of pathology, with 3 representing the most severe case. The panel of three certified medical experts agreed on a severity rating for each sample. Three types of text are found in the recording: (a) continuous speech, (b) three sustained vowels with the onset and offset information, and (c) isolated words, numbers, and common words in Arabic [15].

The text was carefully chosen to include all the phonemes of Arabic. All speakers recorded three utterances of each vowel (/a/,u/, and/i/) to prevent burdening patients, but isolated words and continuous speech were only recorded once. The speech was recorded using the CSL application with a sampling frequency of 44 kHz in the database. Different expert specialists from King Abdul Aziz University Hospital analyzed and validated the vocal problems listed in this database [15]. The AVPD contains 366 samples of normal and diseased patients. Normal participants account for 51% of the total subjects, with the remainder divided among five voice abnormalities (sulcus, nodules, cysts, paralysis, and polyps); 40% are female, while 60% of the subjects are male. All texts were captured and saved in two separate audio formats, namely, WAV and NSP, including three vowels with a repeat, namely, Arabic numbers, Al-Fateha, and common terms. The recorded samples were separated into 22 segments: 6 for vowels (three vowels plus their repetition), 11 for Arabic numerals (zero to ten), and 2 for Al-Fateha (divided in such a way that the first half may be used to train the system and the second part can be used to test the system), and 3 segments for the common word [21].

Many studies use different datasets and classification methods for the detection of various voice pathologies as presented in Table 1 (Figure 3).

Table 1

Similar voice pathology detection and classification method studies

Ref.	Classifier	Dataset	Pathologies
[15]	Kay Pentax CSL Model	MEEI	(1) Unilateral vocal fold paralysis
		SVD	(2) Cyst of vocal fold
		AVPD	(3) Polyp on the vocal folds
[21]	The AVPD has five forms of vocal fold diseases; thus, five classes were presented in this study	AVPD	(1) Sulcus
		MEEI	(2) Nodules
			(3) Cysts
			(4) Paralysis
			(5) Polyps
[22]	GMM	SVD	(1) Vocal polyp
			(2) Sulcus
			(3) Vocal nodules
			(4) Vocal cyst
			(5) Vocal paralysis (unilateral or bilateral)
[23]	SVM	MEEI	(1) Cysts in the vocal folds
		SVD	(2) Vocal fold paralysis on one side
		AVPD	(3) Polyp of the vocal fold
[24]	CNN	SVD	(1) Vocal cyst
		MEEI	(2) Vocal polyp
			(3) Vocal paralysis (unilateral)
[25]	SVM	MEEI	(1) Cysts in the vocal folds
		SVD	(2) Unilateral paralysis of the vocal folds
		AVPD	(3) Polyps on the vocal folds
[26]	GMM	MEEI	(1) Vocal polyp
			(2) Keratosis
			(3) Vocal paralysis
			(4) Nodules in the vocal fold
			(5) Dysphonia
[27]	CNN	SVD	Dysphonia
[28]	CNN	SVD	(1) Paralysis (unilateral)
		MEEI	(2) Polyp
			(3) Cyst of vocal folds
[29]	SVM	SVD	(1) Dysphonia
			(2) Laryngitis
			(3) Recurrent sparse
[30]	DPM	SVD	(1) Reinke’s edema
			(2) Laryngitis
[31]	MFCC	SVD	(AVPD) Paralysis, sulcus, polyp, cyst, nodules
		MEEI	(MEEI) Gastric reflux, hyperfunction, AP squeezing, ventricular compression, and paralysis
		AVPD	(PDA) nodule bilateral, edema pure bilateral, polyp, sulcus
		PDA	(SVD) Hyperfunctional dysphonia, recurrent sparse, laryngitis, functional dysphonia, dysphonia
[32]	CNN, RNN	SVD	(1) Cysts, (2) polyps, (3) nodules, (4) paralysis, (5) sulcus
[33]	OSELM	SVD	(1) Cyst
			(2) Polyp
			(3) Paralysis
[13]	CNN: ResNet50, Xception, and MobileNet	SVD	(1) Cysts, (2) polyps, (3) nodules, (4) paralysis, (5) sulcus

Figure 3

(a) AVPD distribution of healthy and voice-disordered people. (b) The number of male and female samples for each disease and normal subjects [21].

4 Feature extraction techniques of voice pathology

Feature extraction is the calculation of a sequence of feature vectors that provides a compact representation of the input speech signal. This method is generally divided into three components. The first stage is the speech analysis or acoustic front-end, which analyzes the speech signal spectra temporally and creates raw characteristics reflecting the power spectrum envelope of brief speech intervals. The second stage involves the generation of a feature vector that comprises static and dynamic information. The final stage compresses and strengthens the extended feature vectors before handing them to the recognizer [34].

4.1 Mel frequency cepstral coefficients

Mel frequency cepstral coefficients (MFCCs) are the most often used characteristics in the detection of voice pathologies. Using MFCCs to define the human voice generation system is widely accepted. MFCC is a useful method for detecting vocal impairment [5]. MFCC is one of the most commonly used feature extraction methods in speech recognition. MFCC depends on the Mel scale, which is based on the human ear scale; thus, features of the frequency domain, including MFCCs, are more precise than those of the time domain [35,36]. MFCC, which is produced from the fast Fourier transform signal, represents the true cepstral of a windowed short-time signal (FFT). A nonlinear frequency scale is used to simulate the auditory system activity to distinguish MFCC from the true cepstral. Furthermore, these coefficients are stable and reliable considering fluctuations in speakers and recording settings. Simultaneously, MFCC is a technique used to extract features by determining parameters from samples of speech that are similar to those used by humans for hearing speech and simultaneously downplaying all other data [34]. As illustrated in Figure 4, MFCC for various speech samples offers an advantage over other voice characteristics because they can characterize the geometry of the vocal tract completely. An exact representation of the phoneme generated by the vocal tract may be calculated after accurately defining the vocal tract. The envelope of a short-time power spectrum depicts the curvature of the vocal tract, and MFCCs record this envelope properly [5].

Figure 4

The MFCC of normal and pathological voice samples [5].

4.2 Spectrograms

A voice waveform comprises a succession of changing events. This temporal variation relates to spectral properties that drastically shift over time. An STFT is used instead because a single Fourier transform cannot capture this kind of rapid time-varying signal [37]. The STFT, when seen through a sliding window, comprises a separate Fourier transform for each segment of the waveform. The spectrogram can be presented in three dimensions to demonstrate the distribution of power densities with time and frequency, as shown in Figure 5. The figure reveals that the power density distribution of the voice signal substantially changes with time and frequency and may be used to distinguish between normal and sick voices. The graph also demonstrates that the power distribution for normal voice is stable with time and frequency. However, the results for unusual voices are different. The spectrogram is regarded as an effective signal for differentiating between sick and normal voices [5].

Figure 5

The voice spectrograms of normal and diseased voices [5].

5 Voice pathology classification algorithms

One of the most important tasks in voice signal analysis is classifying a specific signal into one of a few predefined categories to arrive at a diagnostic judgment on the vocal problem. The classification of a particular voice signal is quite useful in the diagnostic process. Several classifiers have been used to detect vocal impairment [5]. The primary goal is to distinguish between normal and abnormal speech samples, and the procedure is broken down into three steps: feature extraction, feature selection, and feature classification using a classifier. The audio samples are selected from a voice database, and the required characteristics are extracted. The chosen characteristics are then optimized using an algorithm and sent to the classifier. The classifier divides the speech samples into normal and abnormal categories. Classifiers, such as k-NN and SVM, are commonly used to detect voice dysfluencies. Neural networks and the hidden Markov model (HMM) are among the most popular nonlinear classifiers. SVMs are commonly utilized because they attempt to construct an ideal wall in the feature space that divides the two classes. Two judgments are involved in the two classes. The SVM maximizes the difference in points between the two classes. Linear separation of data is difficult in practice; in such circumstances, SVMs employ the kernel technique to translate the data into a high-order dimensional space and create the dividing wall (hyperplane) in the broad area [1] (Table 2).

6 IoT-based voice pathology

The IoT provides a framework for sensors and devices to connect smoothly inside a smart environment and a simple way to transmit information across platforms. IoT is positioned to become the next breakthrough technology with the recent adaptation of various wireless technologies, thereby maximizing the possibilities of the Internet. The adoption of IOT in smart cities has increased, with an emphasis on developing intelligent systems, such as smart workplaces, smart retail, smart agriculture, smart energy smart transportation, smart healthcare, and smart water [63,64,65]. A large number of communication devices are incorporated with sensor devices in the real world under the IoT paradigm, as indicated in Table 3. Data-gathering devices capture information and send it via embedded communication devices. Various communication systems, including Wi-Fi, GSM, ZigBee, and Bluetooth, are also hired in connecting continuous network devices and objects. These devices broadcast and receive data between remote devices, enabling direct integration with the physical environment to improve living standards via computer-based systems [63]. With the large volume of data and services offered by heterogeneous networks, IoT has envisioned a connected world reality [58,66].

By contrast, cloud computing has risen to prominence, allowing for massive storage and data sharing [58,66]. The combination of cloud and IoT can introduce new and exciting possibilities for both technologies [58,67]. These technologies have the potential to open up a new vista of service sharing, device connectivity, pervasive sensing, as well as enhanced cooperation and communication between people and objects in a dynamic and dispersed performance environment [58,68]. Consequently, various applications are urgently needed in a variety of disciplines within the healthcare industry. Voice pathology is one of the most vital fields in the medical profession. Many people suffer from vocal disorders due to a variety of factors, including severe organ damage, smoking, air pollution, and stress. New studies indicate that more than 7.5% of the total population of the United States suffers from voice pathology. Persons in specific occupations, such as teachers and singers, are particularly affected by voice pathologies because they frequently use their voice [58,69].

Approximately 20% of American teachers have been infected with voice problems [58,69]. Two types of voice pathology detection and assessment are available: objective and subjective. If the method is right, then objective evaluation does not require any extra equipment, and the findings are always unbiased. Meanwhile, subjective evaluations necessitate specialized technology and highly qualified specialists, resulting in a significant expense. Furthermore, subjective assessments vary in accordance with doctors and are dependent on doctor’s knowledge. By contrast, the objective assessment may only be used for initial screening, and the ultimate judgment must be made by medical specialists [58,70].

7 Conclusion

The developments of IoT systems, fog computing, and deep learning methods are currently used in various ways in the healthcare sector. A comprehensive evaluation and taxonomy of the latest research and techniques have been offered in this article to leverage deep learning and IoT in a variety of healthcare applications, primarily in voice pathology identification methods. Furthermore, other issues such as the main feature extraction methods, main datasets used, major classification approaches, research outcomes, key issues, different applications, and recent state-of-the-art applied in voice pathology diagnosis have been emphasized and examined.

The employment of machine learning algorithms in health sectors is predicted to reduce healthcare costs, improve quality of life, and enhance user experience. In addition, these algorithms provide approaches, methods, and tools that may be used to diagnose various speech disorders and will generally help in the improvement of healthcare, specifically with the identification of voice pathologies.

Despite the positive outcomes obtained using machine learning and current technology in healthcare applications, several concerns and unsolved obstacles that require further exploration and discussion, such as the accurate prediction of a large number of parameters, still remain. A large volume of data should be supplied to achieve this goal. Furthermore, understanding diseases and their variants is more challenging than other tasks (e.g., identify the type of voice problem from speech). Therefore, from the standpoint of big data, having a substantial amount of medical data is critical for training strong and effective machine learning models [21]. Healthcare data are extremely complex, incomplete, and unclear. Training a good machine learning model with such a diverse and large set of data is challenging, and several factors such as data sparsity, redundancy, and missing values must be considered.

Ailments change and progress over time. Various models of machine learning are suggested in many sectors of healthcare considering static vector-based inputs. These inputs could not be handled together with the time factor. Thus, a new machine learning approach that can deal with temporal medical data will be an essential component that must be developed. Moreover, creating a new machine learning technique that considers dynamic inputs is necessary, and difficulties in healthcare and biomedicine are becoming increasingly complicated. The diseases are extremely diverse, and most of their reasons remain unknown. In addition, the number of patients in a realistic clinical trial is typically limited; thus, the authors failed to request as many additional patients as needed. A paucity of medical training data is observed for machine learning models. Healthcare organizations deal with multipatient settings where many caregivers perform their duties. In this context, proper identification of caregivers and patients is necessary. Protecting the data gathered by a range of devices and sensors in healthcare against unwanted access is critical. Strict policy and technological security measures must also be adopted to transmit health data to authorized people, corporations, and apps. The creation of an effective algorithm for coordination among protection, detection, and reaction services to avoid various attacks, threats, and vulnerabilities is still in progress.

A future study should investigate extracting enhanced dataset dimensions and traits, such as a new combination of vowel and gender separation. In addition, experimenting with various CNN models and training may help ameliorate methods of speech disorder detection. The allocation of samples with vocal problems is significantly uneven, complicating the diagnosis of voice diseases. A specific type of voice pathology may exist only once in the entire dataset, such as SVD. Therefore, training a specific type of vocal pathology was difficult, which led to low accuracy. Moreover, methods used in speech pathology diagnosis must be highly accurate to prevent any possible errors in the classification procedure. Optimal feature extraction and classifier must be chosen to obtain accurate and effective results of voice pathology detection.