Discriminative Training for Phonetic Recognition of the Holy Quran

Abstract

This paper presents the development of Holy Quran recitation recognizer. The decoder of recognizer performs sub-word level recognition at phoneme. The paper demonstrates high recognition accuracies achieved by applying incremental refinements to the HMM models of the phonemes during the training stage. The Maximumlikelihood (ML) criterion is first applied for HMMs parameter estimation, which produces average recognition accuracies of up to 83 %. This is followed by discriminative technique of minimum phone error (MPE), which is applied to minimize recognition error at phoneme level. Investigation shows that MPE based acoustic models improve generalization. The results show 3-4 % improvement in recognition accuracies, which are promising when compared with the case of ML approach applied alone.

Computational intelligence in processing of speech acoustics: a survey

“…The whole experiment was conducted on Kaldi and achieved good results. Baig et al [24] applied discriminative training criteria, Maximum Likelihood (ML) on Holy Quran followed by MPE. Afify et al [5] introduced an algorithm for simple word decomposition provided with a text corpus and an affix list.…”

Robust Feature Extraction Based on Spectral and Prosodic Features for Classical Arabic Accents Recognition

“…Industrial Revolution: Impact and Readiness Special Issue, 2019 The 39-dimensional features vector of MFCC using 13 MFCC with 13∆ and 13∆∆ features of acceleration coefficients, showed a better result compared to the 25-dimensional features vector, that using 13 MFCC with 12∆ without absolute Energy, and Cepstral Mean Normalization (CMN). The same 39-dimensional features vector of MFCC, also been conducted previously by [12], [13], [15], [16], [18], in their research. The only different is limited to the certain parameter related to the window size of 20ms and 25ms, implemented by [15], [18], respectively.…”

“…The algorithm of spectral features, known as Linear Predictive Coding Coefficient (LPCC) and Mel-Frequency Cepstral Coefficients (MFCC) have been carried out by [11], whereas MFCC only been implemented by [12], [13], [14] for the feature extraction phase. Meanwhile, the Quranic Arabic experiment without accent assessment has been conducted by [15], [16], using the local in-house database constructed by non-Arabs-male certified speakers, and audio in the web for Arabs-male certified speakers, respectively. Both kinds of research also used a conventional MFCC algorithm for the feature extraction stage.…”

“…Furthermore, robust feature extraction is considered a complex issue when dealing with the variability of speech patterns. It can be proven when the result of recognition accuracy is decreased, after implementing Mel-Frequency Cepstral Coefficients (MFCC) or Linear Predictive Coding Coefficient (LPCC) alone as spectral features [13], [15], [17], [18], using the Arabic or Quranic Arabic as data samples. Here, the development of the Quranic Arabic recitation recognizer, conducted using Maximum-likelihood (ML) and Minimum Phone Error (MPE) was presented by [15].…”

“…It can be proven when the result of recognition accuracy is decreased, after implementing Mel-Frequency Cepstral Coefficients (MFCC) or Linear Predictive Coding Coefficient (LPCC) alone as spectral features [13], [15], [17], [18], using the Arabic or Quranic Arabic as data samples. Here, the development of the Quranic Arabic recitation recognizer, conducted using Maximum-likelihood (ML) and Minimum Phone Error (MPE) was presented by [15]. The feature extraction process has been computed using MFCC, which is limited to the Hafs style of pronunciation only.…”