Ali H. Meftah

Comparative evaluation of two arabic speech corpora

The aim of this paper is to conduct a constructive and comparative evaluation between two important Arabic corpora for two different Arabic dialects, namely, Saudi dialect corpus that was collected by King Abdulaziz City for Science and Technology (KACST), and a Levantine Arabic dialect corpus. Levantine dialect is spoken by ordinary Lebanese, Jordanian, Syrian, and Palestinian people. The later one was produced by the Linguistic Data Consortium (LDC). Advantages and disadvantages of these two corpora were presented and discussed. This discussion is aiming to help digital speech processing researchers to figure out the weakness and strength sides of these important corpora before considering them in their experiments. Moreover, this paper can motivate in designing, maintaining, distributing, and upgrading Arabic corpora to help Arabic language speech research communities.

Diacritization, automatic segmentation and labeling for Levantine Arabic speech

It is generally acknowledged that a reliable speech corpus is necessary for any application involving speech processing. In this paper, we propose methods to improve the BBN/AUB DARPA Babylon Levantine Arabic speech corpus to increase its reliability and efficiency. For this purpose, correction of pronunciation, diacritization, and new transcription are performed manually along with automatic phoneme segmentation and labeling. The comparison with the original transcription of the corpus shows a clear improvement in the output results.

Preliminary Arabic speech emotion classification

In this paper, the acoustic features of pitch, intensity, formants, and speech rate are extracted and used to classify the following Arabic speech emotions: neutral, sad, happy, surprised, and angry. Three sentences spoken by four male and four female native Arabic speakers were selected from a newly developed Arabic speech corpus (KSUEmotions). Perception tests using human listeners yielded scores of 87% (male speakers), 84% (female speakers), and 85% (both male and female) accuracy. The best results for the emotion recognition performance were 83%, 56%, and 78% for male, female, and both together, respectively. Anger was the most readily recognized emotion, while happiness was the most challenging to identify. Pitch and intensity features are key in recognizing the Arabic speech emotion of anger.

A Comparative Study of Different Speech Features for Arabic Phonemes Classification

This paper presents the work related to phonetical analysis of classical Arabic speech. Hidden Markov model classifier is applied on Arabic phonemes. For the purpose of this work, a new classical Arabic speech corpus is created. The corpus is based on selected recordings of recitations of The Holy Quran. A number of acoustic features are analyzed and compared. Those are: linear predictive coding (LPC) analysis, mel Frequency cepstral coefficients (MFCC), perceptual linear prediction (PLP), logarithmic mel-filter bank coefficients (FBANK), mel-filter bank coefficients (MELSPEC), and linear prediction reflection coefficients (LPREFC). The confused phonemes and the lowest occurrence rates in many Arabic speech corpora phonemes are investigated. The system obtained maximum accuracies of 85.38% for FBANK feature and 83.37% for MELSPEC feature. The system output results showed that the MFCC and PLP accuracy are too close in accuracy. LPC is not sufficient for automatic speech recognition applications

Revisiting distinctive phonetic features from applied computing perspective: unifying views and analyzing modern Arabic speech varieties

Distinctive phonetic features (DPFs) provide the description of phonemes’ places and manners of articulation. Several, sometimes contradictory, views and definitions of the DPF elements of Modern Standard Arabic have been proposed in the phonology literature. This contrast in views is a significant barrier against utilizing the advantages of DPFs in digital speech processing applications because computer systems do not deliver correct results under vague rules and models. This is a review paper that presents background on Arabic DPFs and in addition to highlighting the historical and geographical verities. It also addresses the problem of ambiguous definitions between classical and modern phonology that may introduce significant challenge to computer scientists and engineers when developing computer systems. Another contribution of this work is to investigate the deviations in phonemes and DPF elements across dialects of Arabic. This is important to provide engineers with better understanding when designing computer software targeting a wide spectrum of Arabic speaking users.

A new look at the automatic mapping between Arabic distinctive phonetic features and acoustic cues

In this paper, the multidimensional phonological feature structure of Arabic is investigated. Our goal is to assess the performance of statistical and connectionist approaches in performing the complex mappings between distinctive phonetic features (DPF) and associated acoustic cues. The present study explores the mapping between 29 phonological voicing, place, and manner features and Mel-frequency acoustic cues. For this purpose, three machine-learning techniques are deployed: Deep Belief Networks (DBN), Multilayer Perceptron (MLP), and Hidden Markov Models (HMM). The three techniques show satisfactory acoustic-phonetic mapping performance and indicate that couple of Arabic DPF elements such as affricatives, alveopalatals, labiodentals, lateral, palatal, pharyngeal, rounded, and uvular have a strong correlation with the acoustic information. The implications of these results on Arabic phonological contrasts are discussed.

Building a Rich Arabic Speech and Language Corpus Based on the Holy Quran

This paper pursues the goal of creating a reliable speech corpus based on The Holy Quran (THQ) audio recordings. Achieving that goal involves major steps to be done and essential requirements to be considered. With the availability of tremendous amount of recordings nowadays, it is of a fundamental importance to select the ones that feature both high audio quality and perfect reciter performance. Also, since the targeted beneficiaries from the corpus are the digital speech processing research community, it is also very essential to maintain an efficient, a familiar and a convenient way of presenting the audio corpus and other language material, such as the language model. Audio recordings of THQ are selected from four sources having a high standard regarding the reciters’ performance. A significant effort is made in phonetical transcription of the audio content such that the written transcript maps perfectly to the uttered phonemes. Furthermore, the corpus dictionary, which is usually required in many fields such as machine learning and datamining, is also created. The first release of the corpus consists of recorded recitations and the necessary metadata of three chapters of THQ of different lengths recited by four reference reciters. Those chapters are selected for this phase based on statistical analysis of the lengths of all chapters and the frequency of occurrence of the Arabic phonemes across all chapters of THQ.

Classical Arabic phoneme contextual analysis using HMM classifiers

This paper presents a phonetic analysis of Arabic speech language phonemes using hidden Markov model classifiers and their confusion matrices. For this purpose, a new classical Arabic speech corpus was planned and designed. The corpus is based on recitations from The Holy Quran of specific scripts. Semi-manual labeling and segmentation of the audio files along with other language resources such as a word dictionary were prepared. Recitations from The Holy Quran are highly indicative of the pronunciation of Arabic phonemes. The classifier results show that phonemes with the lowest frequencies in general have the highest error rates. Overall, the rates of correct classification are 76.04%, 93.01%, 93.59%, and 92.81% for monophone, left and right context biphone, and triphone systems, respectively.

Duration modeling in automatic recited speech recognition

This paper presents a phonetic analysis and recognition of classical and recited Arabic speech phonemes, mainly vowels, using hidden Markov model (HMM) classifiers. For this purpose, a new classical Arabic speech corpus was planned and designed. The corpus is based on recitations extracted from The Holy Quran of specific scripts. For modeling long vowels, we carry out extensive experiments that aim at finding the best way to capture the vowel durations that are semantically relevant in the Arabic language. Our approach consists of adapting the HMM topology to the type of vowels. This method can be applied to other Semitic languages or for the modeling of the geminated phonemes. The proposed method outperforms the baselines systems by achieving an overall correct rate of 87.60% with no specified language model.

Reintroducing KAPD as a Dataset for Machine Learning and Data Mining Applications

KACST Arabic Phonetic Database (KAPD) has been in use by researchers for around fifteen years since its initial release. Researches in acoustics and phonetics have benefited from its phonetically rich content. In fact, KAPD has the potential to go further steps with the research community. In this work, KAPD is subject to enhancements and improvements in order to serve as dataset for machine learning and data mining application. This work involves refining and reviewing the already existing metadata of KAPD and adding new material that are necessary for machine learning and data mining applications. The updated phoneme statistics after the corpus upgrade are presented from different perspectives. Data format and time units are made compatible with those of HTK. The paper discusses the potential of KAPD to serve as either a balanced or an imbalanced dataset.