Accurate Reader Identification for the Arabic Holy Quran Recitations Based on an Enhanced VQ Algorithm

Speech is the most vital interaction tool for human beings of all ages. So, transcribing human speech into words utilizing advances in information technology is challenging because human speech signals vary with different attributes, styles, or environmental noises [1, 2]. The feature of identifying the speaker by utilizing his/her speech for different words with different accents is undoubtedly different from the rules of speech recognition [3]. The difference is that one aims to identify the speaker regardless of the spoken words—the other aims to identify the words regardless of the speaker.

This paper develops an algorithm to identify the reciter of the Holy Quran in Arabic. The Holy Quran was written in Arabic more than 1400 years ago. After that, the Arabic language was developed, and many new accents appeared [4].

Regarding the number of words, Arabic could be considered one of the richest languages. According to the SEBIL center, the Arabic language contains more than 12.3 million words, 20 times more than English [5, 6]. The Arabic language contains 28 characters, where 25 are constant and three are long vowels. These numerous vocabulary counts have made the characteristics of the Arabic language different from any other language. In addition to the richness of Arabic, the Holy Quran recitation has specially formulated rules known as Tajweed [7].

Indeed, Tajweed rules ensure that the recitation is accurate and that each word is pronounced at a moderate speed [8]. Hence, The Holy Quran is considered one of the most important references for the Arabic language. It is also worth mentioning that the study of the Holy Quran is considered a hotspot topic because there are numerous dialects.

In previous studies, researchers focused on extracting the sound features of different recitations of the Holy Quran. They used several methods, i.e., the Mel-Frequency Cepstral Coefficient technique (MFCC). However, they didn’t achieve high accuracy in the extraction process of such systems [9].

Reciting Holy Quran is a duty for every Muslim. Still, this recitation must be free of errors that would conclude readers to wrong interpretations of the verses of the Holy Quran. Thus, Muslims are interested in learning the Holy Quran recitation. This type of learning demands numerous obstacles, for the readers at least. These difficulties include dedicating enough time to learning. First of all, the ultimate goal is to develop an automated Holy Quran recitation system utilizing the latest algorithm in speech recognition. Thus, the target system must analyze the reciter’s voice to decide the reading correctness. Indeed, the reciter’s voice includes features that belong to the reciter’s identity and the read text [10]. Therefore, one of the critical problems is establishing unique speech and language characteristics for each speaker separately.

Moreover, the majority of voice signal processing was limited to the English language. There was scarce work specifically in processing voice signals for the Arabic language and the Holy Quran [11]. Consequently, with all the technological development, these problems must be solved to amplify the number of learners.

This paper proposed a new algorithm for identifying the Holy Quran reciter’s identity. This step utilizes extracted features from the reciter’s voice signal and then detects the identity of this reciter using vector quantization and k-means clustering. Indeed, MFCC was used to extract the feature vectors (FVs) of the speech. Then, a modified version of Linde–Buzo–Gray (LBG) of VQ clustering was utilized to cluster these features into a set of centroids. Moreover, the K-mean clustering algorithm was also used to cluster feature vectors into groups represented by their centroid. Finally, a similarity measure is used to decide the reciter’s identity.

It is worth mentioning that the proposed algorithm of this work should work correctly with all languages, not just Arabic, and all spoken words, not just the words of the Holy Quran. But here, the authors are handling the work with the Holy Quran words, as no significant work is done considering the Arabic language or the Holy Quran’s unique tone.

The remaining of this paper is organized as follows: Section II contains the literature review for the previous related research studies, Section III includes the data preparation process, the MFCC features extraction process, the improved LBG methodology that has been used for generating the codebooks, and the suggested process for matching the codebooks, Section IV presents the measured data, the key findings, the accuracy, and the comparative results,  Section V discusses the obtained results and sum up the conclusions.

Previous research has used several strategies to recognize speakers’ voices, such as those in the studies [20-22]. For instance, the familiarity and voice recognition for acoustic-based representation to find the voice averages [20]. Moreover, the voice principle with non-native language speakers has been investigated [21]. Moreover, an enhanced approach to children’s voices is utilized by Arunachalam [22]. This is to recognize the speech with hearing impairment by multi-sets of features and models.

However, as discussed before, the Holy Quran reciters follow stringent rules on top of the complexity of the Arabic language. Therefore, in this section, the process of the proposed method to recognize the Holy Quran reciter is presented. Figure 1 shows a flowchart of the proposed Holy Quran reciter identification system, including four stages, which are: the data preparation stage, MFCC feature extraction stage, codebook matching stage, and codebook matching stage. Each one of these four stages will be discussed in the following sections:

1.png

Figure 1. Flowchart of the proposed Holy Quran reciter recognition system

3.1 Data preparation and speech signals segmentation

Input speech signals usually have noise and zones of unwanted silence. Hence, in this paper, the method introduced by Giannakopoulos [23] is implemented using MATLAB^®. In this paper, the proposed system’s software was built and executed using MATLAB®. Hourri and Kharroubi algorithm removes the defect areas and produces convenient voice segments. Moreover, it comprises the signal energy, the spectral centroid methods, and the simple threshold criterion. Thus, by applying it, the Holy Quran verses, words, or even the letters’ sounds could be easily separated and filtered probably. The filtration process can be done by changing the length of the non-overlapping short-term windows (frames) and the step’s length. Furthermore, a min-max normalization method defined in Eq. (1) has been applied to the speech signal to ensure that the signal amplitudes of the different reciters are in the same range.

$Signal _{n o r m}=\frac{S-S_{\min }}{S_{\max }-S_{\min }}$      (1)

2a.png

Figure 2. Surah Al-Kawthar’s signal

where, S is the input signal, S_min is the minimum amplitude of all the signals in the database, and S_max is the maximum amplitude. Figure 2 shows the recorded signal for “سورة الكوثر / Surah Al-Kawthar” from the Holy Quran by a professional reciter after normalizing its amplitude.

It can be seen that the verses have been separated correctly, and all noises have been filtered. In this figure, the brown signals are the speech signal, and the gray ones are the areas of silence between verses.

3.2 MFCC features extraction

Feature extraction is a process in which speech signals are analyzed. The main features of the voice get extracted to be used in an automatic speech recognition system [8]. The MFCC is widely used for extracting voice features, and it is comprised of coefficient vectors related to each frame of voice [24]. Ahmad et al. [25] mentioned that the MFCC approach is the best method for analyzing Holy Quran verses, and it gives the highest accuracy. The process of implementing MFCC features can be briefly described in the block diagram shown in Figure 3 [8].

2.png

Figure 3. MFCC voice signal features extraction process

In addition, the formula that has been used to compute the MFCC features is written in Eq. (2) [24].

$\operatorname{Mel}(f)=2595 \times \log _{10}\left(1+\frac{f}{700}\right)$      (2)

Several Feature Vectors represent the MFCC feature coefficients (FVs); each has the same number of coefficients but with different ranges. In addition, a log-energy FV (the zeroth coefficient) is often excluded because it carries only a tiny amount of information about the speaker [26, 27].

This paper utilizes a set of 12 MFCC FV coefficients, excluding the log-energy coefficients. The characteristics of the implemented MFCC function include the windowing length, which describes the length of each word segment or frame, which in this case is equal to 23 mS for each frame. This window length value was found suitable, as suggested in [2]. Furthermore, these frames are overlapped to ensure no information is lost at the end of each frame. Therefore, the overlap length equals 11.5 mS, and every frame’s data exists in its previous and next frames’ data. These two characteristics have an inverse relationship with the number of MFCC coefficients extracted for each speech signal. Increasing the time of the window length will decrease the number of MFCC coefficients and vice versa.

3.3 MFCC features clustering

Clustering is realized as finding homogeneous groups of data points in a dataset [28]. Each set of these groups is called a cluster, and each cluster has a center point named a centroid. Thereby, feature clustering algorithms can be used effectively for feature-matching techniques. Many clustering algorithms have been used recently for speaker recognition, including Vector Quantization (VQ), K-means Clustering, Support Vector Machine (SVM) [29], Neural Network (NN) [30, 31], Hidden Markov Modeling (HMM), and Dynamic Time Warping (DTW) [32]. In this paper, the VQ and K-means tactics have been used to identify the clusters. For feature clustering, this work proposes an improved VQ clustering algorithm based on the LBG-VQ technique in addition to utilizing k-means clustering for accuracy validation and comparison purposes.

3.4 Vector Quantization (VQ) algorithm

VQ is a data compression technique in which the data points are squeezed into a smaller dataset. Data compression is achieved by remapping these points to a finite even number of clusters represented by their centroids [29]. A code vector is defined as the centroid value of the cluster obtained by VQ. Each code vector contains the key points that represent all data points. The collection of all code vectors in a dataset is called the codebook [33]. A simplified two-dimensional VQ diagram shown in Figure 4 illustrates the conceptual Vector quantization. This figure shows the datapoints features of a Holy Quran reciter (a data point refers to one MFCC coefficient in two-dimension FVs). The clusters’ data points are represented as colored dots, and their centroids are represented as crosses with the same color. The VQ technique refers to the centroids of the clusters as a code vector. For this reason, the VQ is considered a data compression technique [32].

4.png

Figure 4. Simple VQ diagram

Clustering technique performance could be measured using clustering error criteria, the distance of every data point in a cluster from its centroid, which should be as lowest as possible [31]. Several algorithms were used to compute the clustering error criterion. However, this paper utilizes the squared Euclidean distance (SED), as described in equation (3). To decide on a point to join a cluster, the SED should always be the lowest to join this point to any other cluster. The SED for two vectors, V1 and V2, is described in formula (3) where N is the dimension of V1 and V2.

$\operatorname{SED}\left(V_1, V_2\right)=\sqrt[2]{\sum_{i=0}^{N-1}\left(V_{1 i}-V_{2 i}\right)^2}$       (3)

3.5 Proposed clustering algorithm based on LBG-VQ

After the FVs have been computed from a reciter’s segmented input speech using MFCC, the codebook is created based on clustering the FVs, called a reciter-specific VQ codebook. The proposed clustering algorithm developed based on the BG-VQ algorithm clusters a set of L FVs to M (8 or 16) codebook vectors, where M is less than L [32, 33]. The proposed algorithm and the implementation are formally described in the following recursive procedures:

1. The input is a set of L FVs. Each FV has 12 elements, as described in the MFCC feature extraction section. Initially, all FVs are grouped in one cluster.
2. The centroid of the initial cluster, the mean, is calculated as described in Eq. (4). Then this centroid is stored in the M1 codebook. The dimension of the obtained centroid vector is 12.

$M_1(L)=\frac{\sum_{i=1}^k F V_i(L)}{k}$      (4)

3. Each centroid, which represents one cluster of data, is utilized to generate two new centroids by adding and subtracting a minimal value, namely Epsilon (e). The value of e depends on the elements’ values in each FV and is computed as proposed in Eq. (5).

$\varepsilon(\mathrm{L})=(\max (\mathrm{L})-\min (\mathrm{L})) * \mathrm{~W}$       (5)

where, W is a weight factor. This study found that the best value for W equals 15%. Accordingly, M₁ and M₂ can be computed in Eq. (6).

$M_{1,2}(L)=M_{1,2}(L) \pm \varepsilon$       (6)

4. Afterward, the original cluster is divided into two clusters. Each FV in the original cluster is added to a cluster such that this FV is closer to the centroid of the joined cluster. To achieve this, D₁ and D₂ are computed according to (7).

$D_{1,2}=\sqrt[2]{\sum_{i=1}^k\left(F_i(L)-M_{1,2}(L)\right)^2}$      (7)

5. After clustering all the FVs, the contoids M₁ and M₂ for the two clusters are recomputed. If the new centroids are not equal to the old ones, then steps 3 and 4 are repeated.
6. Hence, two clusters with two centroids will be formed. Likewise, the same steps are repeated to have four centroids, eight centroids, and 16 centroids.
7. Finally, a codebook with size N centroids can be created. Figure 5 illustrates the detailed flow chart steps of this improved LBG algorithm [32].

3.6 Codebooks matching

Codebook matching is defined as the process of searching in the reciters’ database to identify the current reciter. First, one retrieves a codebook from the reciters’ database and matches it with generated codebook. The generated codebook for the Holy Quran reciter is compared with all codebooks stored in the reciters’ database. Each centroid in the generated codebook is coupled with its nearest centroid in the retrieved codebook. These two codebooks would be from different verses of similar or different reciters. The purpose of this step is to measure the matching error between the two codebooks based on the total SED between the coupled centroids. So, if the total SED is less than the threshold criteria, the reciter of the retrieved codebooks would be considered a candidate reciter for the generated one and added to a reciters’ candidate list. After comparing the generated codebook with all codebooks in the database, the reciter with the lowest matching error in the reciters’ candidate list is considered the reciter for the input signal. If the reciters’ candidate is empty, then the reciter of the input signal is considered unknown.

For matching reciter-generated codebook (RGCB) with database codebooks, the following procedure has been performed:

1. Read the codebook from the codebook database (CDB).
2. Compute the SED between the centroids of RGCB and CDB and build a comparison matrix as shown in Table 1.
3. Search the CED to find the smallest value, couple the row centroid with the column centroid, and add them to the centroid coupling array as shown in Table 2. Then, delete the selected row and the selected column from the comparison matrix.
4. Repeat step (3) until the comparison matrix is empty.
5. Repeat step (2) until the N centroids of the 1^st codebook match with the N centroids of the 2^nd codebook.
6. Further steps are added to ensure that no two or more centroids of the 1^st codebook match with the same centroid of the 2^nd codebook.
7. Calculate the total SEDs, the sum of all the SED in the coupling array.
8. If the total SEDs are less than a threshold, add this CDB to the candidate reciters array.
9. Repeat all these steps for each codebook in the codebook database. 
10. If the candidates’ reciter array is not empty, the recognized reciter is the reciter with the least total SEDs; otherwise, the reciter is not identified as one in the codebook database.
11. Calculate the total SED, the sum of all the SED in the coupling array.

Table 1. Calculated SED between RGCB and CDB

Table 2. Coupled RGCB and CDB centroids and their SED

For validating purposes and ensuring that the matching procedure is accurate and reliable, The total SED results from matching the 1^st codebook and the 2^nd codebook of the same verse of different reciters must have the same value when doing this interchangeably. Additionally, to evaluate the proposed algorithm, the recognition rate is defined as the percent of the correctly recognized reciter with respect to the total conducted recognitions. The recognition accuracy rate (RAR) is formulated as in formula (8), which is simply as the known error rate formula [34].

$R A R=\frac{\text { No. of correct recogntion }}{\text { Total No. of recognition experments }} * 100 \%$      (8)

Furthermore, the proposed algorithm performance evaluation metrics (precision and recall) were computed in Eqns. (9) to (10) [35].

Precision $=\frac{T P}{T P+F P}$     (9)

Recall $=\frac{T P}{T P+F N}$      (10)

where, “TP, TN, FP, and FN are the True Positive, True Negative, False Positive, and False Negative values, respectively.

5.png

Figure 5. Flowchart for the proposed clustering algorithm

The reciter that needs to be identified should have the minimum total SED value among all the other values. In other words, the matching error criterion for the same reciter based on different verses is expected to obtain the lowest value. Therefore, the above threshold value has been selected to achieve the best reciter identification accuracy. Table 6 illustrates the total SED values for five reciters. Each row shows the comparison between reciters reciting the second verse for “سورة الكوثر / Surah Al-Kawthar” with reciters reciting the first verse.

Additionally, this table shows sample results of whether a reciter has been identified. The sample result gives a recognition rate of 100% by setting a threshold value of less than 20, as illustrated in Table 6. Obviously, a zero SED value will be obtained if the same verse is used for comparison.

To evaluate the proposed reciter identification system, a codebook database has been developed for fifteen professional Holy Quran reciters where five of them are well-known Holy Quran reciters. The codebook (feature vectors) is computed for each verse and stored in the codebook’s database. In the first phase, the recitations for the fifteen reciters were recorded, where each one of them recited the four verses of “Surah Al-Kawthar.” In the second phase, some reciters are selected to be unknown. Then, they were asked to recite the first verse.

Table 6. Total SEDs comparison for reciter utilizing first and second verse of Surah Al-Kawthar

The proposed reciter identification system is utilized to identify the unknown reciters, and the recognition accuracy rate (RAR) is computed. Table 7 shows RAR for identifying the reciter where each row shows the result when utilizing one verse as input. The overall reciter identification accuracy equals 98.21%.

Table 7. The overall accuracy rate of the system

For further performance study, precision and recall are also computed in the proposed system. This study presented the accuracy of the proposed algorithm utilizing 8 to 16 centroids for the codebook. Table 8 depicts the precision and recall for the prosed system utilizing 8 and 16 centroids codebooks. They have been computed based on each verse in Surah Al-Kawthar as input. Verse #2 shows the best precision and recall for identifying the reciter. The overall precision equals 95% and recall equals 96%.

Table 8. Precision and recall for the proposed algorithm utilizing 8 and 16 centroids

Furthermore, a comparison study has been conducted with recent related research. Table 9 presents the comparison with the previous studies’ total accuracy results. It can be noticed that the proposed system achieved the highest accuracy percent with 98.21% using the proposed system based on an improved VQ algorithm utilizing 16 centroids for the codebook (feature vector).

Table 9. Total systems’ accuracy results compared with the literature