Quran reciter identification using NASNetLarge

Speaker identification is a technique that uses features extracted from speech signals to automatically figure out who is speaking. Due to its potential usefulness in a variety of everyday activities like surveillance applications, biometrics applications, business transactions, telephone banking, access control systems, law enforcement, and voice mail surfing, it has recently gained significant scientific interest. The difficult work of speaker identification starts with gathering the speaker's audio data, followed by extracting audio features that can be utilized to identify each speaker. Finally, selecting the appropriate model to train it with the extracted features to achieve correct identification [1]. Hence, feature extraction methods and classification algorithms are essential parts of a speaker identification system [2]. Various tools and techniques with several rates of success have been created for speaker identification systems. Features can be implemented as Zero Crossing Rate (ZCR) [3], Mel-Frequency Cepstral Coefficients (MFCCs) [4], Spectral Centroid (SC) [5], pitch frequency [6], etc. MFCC features are the most effective and popular features for acoustic modelling, particularly for speaker identification [7]. For speaker identification task, machine learning models such as Linear Regression [8], Support Vector Machine (SVM) [9], and k-Nearest Neighbour (KNN) [10] were implemented. However, the efficacy of such machine learning models is reliant on the richness and variety of their features [11]. Also, deep learning models have been successfully employed in speaker identification tasks, even though there are some concerns with deep learning techniques. Whereas, starting from the beginning with a fresh model for deep learning necessitates an enormous amount of data, long periods of training, and vast computational power. Consequently, a transfer learning approach can indeed be used to get over the concerns of the traditional deep learning model [12]. The deep transfer learning technique can be applied by using the new training data to update the weights for a previously trained model on a new task. Another way to apply deep transfer learning models is to implement a previously trained model as a feature extractor, followed by a classifier or simple model to get identification [13]. Deep transfer learning has many benefits, such as less time spent training, improved neural network efficiency, and a lot of data is not necessary.

Despite many studies have been done on speaker identification, relatively few of them have specifically focused on the Holy Quran reciter identification. For Muslims, nothing is more sacred than the Holy Quran. Muslims perform five daily prayers, and during their prayers, they must read from the Holy Quran. Many Muslims, regardless of language, listen to and read the Holy Quran, especially during Ramadan (the Islamic month of fasting). Muslims cannot always identify the reciter of the Holy Quran when they hear it. Since the Holy Quran is recited using a unique mechanism, certain guidelines must be followed at all times, known as Tajweed [14].

In this paper, an identification model for a Holy Quran reciter is provided. Six pre-trained deep learning models are assessed after being separately integrated into our proposed model. A Quranic dataset for twenty reciters of the Holy Quran is created based on reprocessing audio files to an image-based visual representation depending on MFCC features from about 11,000 audio files. This paper is the first to use deep transfer learning models for efficient application in Quranic reciter identification.

The remaining sections are organized as follows: The related work on the Arabic language, especially for the Holy Quran, is in Sect 2. Section 3 presents a background on MFCC features, in addition to deep transfer learning approaches and its pre-trained models. Section 4 describes the dataset that was applied to the proposed model. Section 5 explains the proposed model for identifying the reciter of the Holy Quran. Section 6 talks about the results and discussion of the experiment depending on the rate of recognition. Section 7 concludes this paper and talks about future work.

Many researchers have explored speech processing; despite that, the majority of their efforts are concentrated on English, and few of them have particularly targeted Arabic. Although there are many challenges with Arabic speech that can be highlighted, particularly with the Holy Quran. Reciting the Holy Quran correctly is one of these challenges, whereas when reading the Holy Quran, some unique procedures and guidelines must be followed, known as Tajweed rules. Ayyoub et al. [15] attempted to solve the problem of finding the right way to use Tajweed rules throughout the Holy Quran. In particular, they looked at eight Tajweed rules that beginners in recitation have to deal with. They combined traditional features (such as Linear predictive Code (LPC) [16], MFCCs [4], Hidden Markov Model-based Spectral Peak Location (HMM-SPL) [17], and Wavelet Packet Decomposition (WPD) [18]) with those retrieved by a convolutional Deep Belief Network [19] and utilized SVM for classification. For an internal dataset of thousands of audio recordings, the obtained accuracy was 97.7%. Alagrami and ljazzar [20] also conducted automatic recognition of the Arabic recitation rules of the Holy Quran (smart Tajweed). Alagrami and ljazzar [20] used filter banks [21] to extract features as a baseline method, and for classification, they used SVM. This model achieved 99% validation accuracy for only four Tajweed rules. In addition to Tajweed, Makhraj (the parts of the mouth from which Arabic alphabets are uttered) is an important thing Muslims should know to correctly read the Holy Quran [22]. As a result, distinguishing the Makhraj from the reciter is another challenge of the Holy Quran. Hamid et al. [23] developed an approach for Makhraj recognition employing MFCC features and Mean Squared Error (MSE) for pattern matching of the hijaiyah letter. This model achieved 100% precision in a dataset that was created for people between the ages of 21 and 23 who are experts in Makhraj utterance. Also, the categorization of specific melodies, known as maqams in Arabic, by the Holy Quran's reciters is a further instance of the Holy Quran's challenges. Shahriar and Tariq [14] implemented MFCC features and Artificial Neural Network (ANN) that consisted of five deep layers to classify the maqams of the Holy Quran recitations. Shahriar and Tariq [14] achieved a 95.7% accuracy rate on a dataset created from two Quranic reciters. Identifying the reciter of the Holy Quran is also a difficult challenge. Although there exists a unique signal for each Quran reciter, these individual signals tend to converge due to the influence of Tajweed rules and recitation style. Alkhateeb [24] developed an algorithm using machine learning for the identification of Holy Quran reciters. MFCC features are extracted from ten reciters, and for classification purposes, the KNN classifier and the ANN classifier are both used. The ANN achieved 97.62% accuracy for Chapter 18 and 96.7% accuracy for Chapter 36, while KNN achieved 97.03% accuracy for chapter 18 and 96.08% accuracy for chapter 36. Also, Anazi and Shahin [25] used the same model for reciter identification as in a related study [24]. But in [25], the authors used another 10 reciters and different Quranic chapters to construct the MFCC features data set. ANN performed with an average accuracy of 98.5% in Chapter 7 and 97.2% in Chapter 32, while KNN's average accuracy for Chapter 7 is 97.02%, compared to 96.07% for Chapter 32. Nahar et al. [26] suggested two models: ANN and SVM to identify the Quranic reciter out of 15 reciters using MFCC features. The accuracy rate reached was 96.59% using SVM classifier and 86.1% using an ANN. This model is applied to a dataset composed of 230 verses for each of the 15 reciters. In addition, Shah and Ahsan [27] introduced an Arabic speaker identification system using the combination of Discrete Wavelet Transform (DWT) [28] and LPC features [16]. Shah and Ahsan [27] achieved a 90.90% recognition accuracy in the dataset for five reciters.

This section describes the necessary background information, which includes the MFCCs audio feature extraction step as well as pre-trained models for deep transfer learning.

The Holy Quran is divided into 30 chapters and contains 114 surahs. Each chapter has a distinct number of surahs, and each surah has a variable length. Long verses can be found in the surahs Al-Bakara and l-Emran, which have 286 and 200 verses, respectively. However, other surahs, such as Alikhlas, Alfalaq, and Alnas, include only a few verses 4, 5, and 6, respectively. This study uses 7 surahs and the audio recitations of twenty different reciters as its dataset. To make the identification process more challenging, the reciters' voices were chosen to be as similar as possible. The MP3 recitations of the Quran were found in a publicly accessible audio dataset at http://​ourquraan.​com/​, from which the collection was obtained. The length of each clipped audio sample is exactly 20 s. Overall, the database is composed of 11,000 audio segments, with 550 segments allocated for each reciter. The names of different surahs of the Quran are listed in Table 1.

In addition, the names of twenty distinct reciters are included in Table 2.

This proposed model focuses on determining the reciter of the Holy Quran; its pipeline is shown in Fig. 9. In this work, the voice input is made up of various audio files containing the reciter's voice signals, whereas each audio file consists of 20 seconds. Since this work, consider the speaker identification task as if it is an image classification problem. A pre-processing phase is applied first to convert audio files from the sound representation to the image representation before moving on to the extraction phase. Therefore, the MFCCs are obtained from audio samples to apply the pre-processing phase. To get the MFCCs [7] first the audio signal is pre-emphasized to boost the amount of energy in the high frequencies and make them more available to the acoustic Model [30]. The pre-emphasis is applied to our audio files using a transformation filter, which is given by Eq. 1.where \({S}_{n}\) is the input signal and \({S}^{\prime}_{n}\) is the output after applying pre-emphasis.

After that, the incoming pre-emphasized audio signal is segmented into frames. In this study, each frame is 25 milliseconds long, and the frames next to it overlap by 10 milliseconds. After the data are split into frames, the window function is used to multiply each frame as declared in Eq. 2.where \(N\) is the quantity of samples, \(W\left(n\right)\) is the window applied to the original input sample, \({S}^{\prime}_{n}\) and is the pre-emphasized audio signal.

The Hamming window is used for windowing [44], which is typically expressed as Eq. 3.

Discrete Fourier Transform (DFT) [45] is then performed on each frame to transform the signal from its time domain representation to its frequency representation using Eq. 4 in order to get the speech signal's spectrum.where \(X\left(k\right)\) is the frequency representation that matches up its corresponding time domain signal \(X\left(n\right)\).

Following that, Mel-scale power spectrum coefficients from the audio frequency representation are calculated by applying the following two steps. Firstly, the Mel frequencies for the lowest and highest linear frequencies are calculated using Eq.uation 5.where \(Mel\left(f\right)\) is the Mel frequency that matches up with the linear frequency \(f\).

Secondly, a triangular Mel-scale filter bank [21] is applied to highlight significant frequencies in order to mimic what a human would perceive. Mel-scale power spectrum coefficients are calculated using Eq. 6.where \(M\) is the Mel filter bank number, \(K\) is the DFT bin number, \(X\left(k\right)\) is the DFT point of the particular window frame of the input signal, and \(W\left(K\right)\) is the triangular Mel-scale filter.

Since the Mel-spectral values for each frame are closely correlated, it is necessary to apply an appropriate transformation. In this work, the Discrete Cosine Transform (DCT) is used to produce uncorrelated cepstral features using Eq. 7.where \(S\left(m\right)\) is the output of the filter bank and \({C}_{n}\) are the cepstral coefficients.

Until now, 12 cepstral coefficients per frame were extracted. The 13th coefficient energy [30] can also be calculated in each frame using Eq. 8 from the time domain signal without any windowing.where \(x(n)\) is the audio, and N is the frame’s length.

After completing the calculations, a total of 13 MFCC coefficients have been extracted per frame. The proposed method uses only these 13 MFCC cepstral features per frame to reduce the dimensionality of feature space and make the model feasible for real-time implementation. Those steps are applied to all frames on the whole audio file until MFCC features sequences are obtained for the audio file. Following that, MFCC features sequences are visualized to get the MFCCs images, which represent the visual representation for the audio file. The previous steps applied to 11,000 reciters audio files. Figure 10 represents one of the MFCC features sequences representation for one of the twenty reciters of the Holy Quran.

After that, the second approach of deep transfer learning which uses the pre-trained model as a feature extractor is used. The main advantage of the second approach is that the data are only run once on the pre-trained model, rather than once every training epoch. So, it is much faster and less computationally expensive. Six pre-trained models are adopted in this stage separately: NASNetMobile, NASNetLarge, EfficientNetB7, EfficientNetV2S, EfficientNetV2M, and EfficientNetV2L. Each MFCCs image that represents an audio file is resized to match the size of the pre-trained model. The size of the image for each pre-trained model is shown in Table 3.

The features extracted by the pre-trained model are inputs to a flattened layer, which flattens the pre-trained model's output data. Following that, two dense (1024) hidden layers are applied. For both hidden layers, a Batch Normalization (BN) [46] layer is applied between the linear layer and a nonlinear activation function ReLU [47]. BN is performed to avoid overfitting [48] and normalize the input data for the next layer so as to combat the internal covariate shift [46] issue, in addition to improving the learning speed of neural networks. BN is applied to all batches, whereas each batch has \(m\) samples. The initial step of BN is to calculate the inference mean \({E}_{x}\) and the inference variance \({Var}_{x}\) of the layer inputs by Eqs. 9 and 10, respectively [46].where \(j\) is the number of batches, and batch mean \({\mu }_{B}\) as well as batch variance \({\sigma }_{B}^{2}\) are obtained using Eqs. 11, and 12, respectively [46].

After that, \({{\varvec{E}}}_{{\varvec{x}}}\) and \({{\varvec{V}}{\varvec{a}}{\varvec{r}}}_{{\varvec{x}}}\) are utilized to normalize the layer inputs as well as \({\varvec{\delta}}\) and \({\varvec{\beta}}\) are used to scale and shift the normalized input to obtain the layer's output. The normalized data are given by Eq. 13.where \({\varvec{\delta}}\) and \({\varvec{\beta}}\) are learning parameters of BN.

Also, the mathematical representation for the ReLU activation function which is fed by the normalized data is given by Eq. 14.where \(x\) is the normalized data. Finally, the outputs of the two layers are fed into a SoftMax activation function [49] to identify the twenty identity classes for Holy Quranic reciters. The SoftMax function is adapted using Eq. 15. It generates outputs that are a range of values between 0 and 1, with total probabilities equal to 1, whereas the highest probability represents the class of the Quranic reciter.

In this section, experiments on our created dataset are applied to evaluate the proposed models. The experimental settings, training details, and performance metrics are presented before the results are shown.

This research introduces a proposed dataset for twenty reciters of the Holy Quran, comprising 11,000 audio files that have been converted from an audio representation to a visual representation using the Mel-Frequency Cepstrum Coefficients. Furthermore, we propose the first Holy Quran reciter identification model that applies a transfer learning approach. Six pre-trained deep learning models, including NASNetMobile, NASNetLarge, EfficientNetB7, EfficientNetV2S, EfficientNetV2M, and EfficientNetV2L, were assessed after each was integrated separately as a feature extractor in our proposed model. On evaluating the proposed models on the suggested dataset, the NASNetLarge model achieved the highest accuracy of 98.50% when used for feature extraction.

Future plans include increasing the number of Holy Quran reciters and improving identification accuracy by adopting newer extraction methods and classification algorithms. Since the Holy Quran can be recited in ten different styles, determining the recitation style will also pose a challenge in the future.