The objective of this paper is to study the acoustic word stress in Arabic. We performed an acous... more The objective of this paper is to study the acoustic word stress in Arabic. We performed an acoustic analysis based on the three prosody parameters (fundamental frequency, intensity and duration), for a corpus of 100 sentences (about 400 words) in order to detect the stressed syllables. This study is based on our previous syllabification paper, and will help in determining the stress assignment rules for our Arabic Semi-Syllable Synthesizer. We discuss stress assignments on syllables of monosyllabic, disyllabic and polysyllabic words in addition to phrases with liaison cases. The assigned stress can have one of three degrees: primary stress, secondary stress and neutralized (absence of stress). In this paper, we compared the acoustic characteristics of stressed syllables with non-stressed ones. The comparison showed that the duration cue has the most prominent effect followed by intensity. F0 is the less effective cue. The primary stress is assigned to the syllable having long vowel...
Speech enhancement is one of the challenging tasks in signal processing, especially in the case o... more Speech enhancement is one of the challenging tasks in signal processing, especially in the case of non-stationary speech-like noise. In this paper we propose a new supervised speech enhancement system that uses Fischer Discriminative Dictionary Learning (FDDL) algorithm to model both speech and noise amplitude spectrum, where the cost function accounts for both "source confusion" and "source distortion" errors. In the enhancement stage, we use sparse coding on the learnt dictionary to find an estimate for both the clean speech and noise amplitude spectrum. In the final stage, a Wiener filter is used to refine the clean speech estimate. Experiments on NOIZEUS dataset using two objective speech enhancement measures: frequency-weighted segmental SNR and Perceptual Evaluation of Speech Quality (PESQ) demonstrate that FDDL outperforms other tested dictionary learning algorithms in the presence of considerable noise (0 dB) for all studied noise types, and in the presen...
Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that c... more Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that could be words, syllables or phonemes. Phonetic segmentation is about finding the exact boundaries for the different phonemes that composes a specific speech signal. This problem is crucial for many applications, i.e. automatic speech recognition (ASR). In this paper we propose a new model-based text independent phonetic segmentation method based on wavelet packet speech parametrization features and using the sparse representation classifier (SRC). Experiments were performed on two datasets, the first is an English one derived from TIMIT corpus, while the second is an Arabic one derived from the Arabic speech corpus. Results showed that the proposed wavelet packet de composition features outperform the MFCC features in speech segmentation task, in terms of both F1-score and R-measure on both datasets. Results also indicate that the SRC gives higher hit rate than the famous k-Nearest Neigh...
Journal of Telecommunictions and Information Technology
Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that c... more Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that could be words, syllables or phonemes. Phonetic segmentation is about finding the exact boundaries for the different phonemes that composes a specific speech signal. This problem is crucial for many applications, i.e. automatic speech recognition (ASR). In this paper we propose a new model-based text independent phonetic segmentation method based on wavelet packet speech parametrization features and using the sparse representation classifier (SRC). Experiments were performed on two datasets, the first is an English one derived from TIMIT corpus, while the second is an Arabic one derived from the Arabic speech corpus. Results showed that the proposed wavelet packet decomposition features outperform the MFCC features in speech segmentation task, in terms of both F1-score and R-measure on both datasets. Results also indicate that the SRC gives higher hit rate than the famous k-Nearest Neighbors (k-NN) classifier on TIMIT dataset.
Received: 18 July 2020 Accepted: 6 November 2020 In this paper, we present a hybrid optimization ... more Received: 18 July 2020 Accepted: 6 November 2020 In this paper, we present a hybrid optimization framework for gridless sparse Direction of Arrival (DoA) estimation under the consideration of heteroscedastic noise scenarios. The key idea of the proposed framework is to combine global and local minima search techniques that offer a sparser optimizer with boosted immunity to noise variation. In particular, we enforce sparsity by means of reformulating the Atomic Norm Minimization (ANM) problem through applying the nonconvex Schatten-p quasi-norm (0<p<1) relaxation. In addition, to enhance the adaptability of the relaxed ANM in more practical noise scenarios, it is combined with a covariance fitting (CF) criterion resulting in a locally convergent reweighted iterative approach. This combination forms a hybrid optimization framework and offers the advantages of both optimization approaches while balancing their drawbacks. Numerical simulations are performed taking into account the...
Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and sho... more Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and short vowels which are written as punctuation marks, above or below letters. Those short vowels are normally omitted in Arabic texts because the reader can fill them and guess the meaning based on his knowledge of the language, and the context in which the words are. However, with the widespread usage of computers in linguistics application; Arabic texts need to be supplied with short vowels in order to be analyzed. Search engines, text to speech engines, and text mining tools are just some examples of applications that need Arabic texts to be vocalized before being processed. In this paper, we present a new method to supply those vocals. The approach emphasizes on unsupervised machine methods, because public Arabic corpora are not available. Arabic rich morphology and diverse orthography present serious challenges for this approach. An algorithm is developed and a system is implemented in ...
Arabic Language Processing is gaining increasing importance all around the world. This language i... more Arabic Language Processing is gaining increasing importance all around the world. This language is spoken by nearly 300 millions in the Arab World, and is an interesting language for the 1.3 billion Muslims. Arabic is becoming also a focal point of interest in many universities all around the world. The positive aspect of this interest in Arabic -which was relatively ignored when compared with other live languages-, is the intensive work on language technologies, and an increasing amount of digital contents on the Internet. Arabic Language is one of the most promoted research axis in HIAST since its foundation in 1983 (Ghneim & Al-DAkkak, 2006). In the introduction of this paper, we mention the most relevant works, in both speech and text aspects, which can be of interest in many applications, and can be a subject of resource sharing. In the second section we detail our acoustic database of semi-syllables and in the third one we develop our morphological analyzer.
This paper presents a rule-based approach to detect Arabic Speech Act types. Our approach is base... more This paper presents a rule-based approach to detect Arabic Speech Act types. Our approach is based on the fact that Arabic language is rich in morphology and that its syntax is inherently rule–based. The importance of Arabic Speech Act type recognition comes from its wide variety of application domains, such as: speech synthesis, speech and emotion recognition and conversational analysis. In this work, a rule-based Expert System has been developed in a bootstrapping manner, to classify an utterance speech act type to one of the sixteen types (interrogation, exclamation, negation, etc.). Our system was tested on a corpus of about 1500 sentences and achieved an accuracy rate of 98.92%. The expert system was used to automatically annotate a corpus (sentence-level annotation) by assigning a tag to each sentence according to its speech act type.
Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and sho... more Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and short vowels which are written as punctuation marks, above or below letters. Those short vowels are normally omitted in Arabic texts because the reader can fill them and guess the meaning based on his knowledge of the language, and the context in which the words are. However, with the widespread usage of computers in linguistics application; Arabic texts need to be supplied with short vowels in order to be analyzed. Search engines, text to speech engines, and text mining tools are just some examples of applications that need Arabic texts to be vocalized before being processed. In this paper, we present a new method to supply those vocals. The approach emphasizes on unsupervised machine methods, because public Arabic corpora are not available. Arabic rich morphology and diverse orthography present serious challenges for this approach. An algorithm is developed and a system is implemented in Java. The techniques presented in this dissertation can be applied to similar Semitic or other languages that have the same problem.
In this paper, we reformulate the gridless direction of arrival (DoA) estimation problem in a nov... more In this paper, we reformulate the gridless direction of arrival (DoA) estimation problem in a novel reweighted covariance fitting (CF) method. The proposed method promotes joint sparsity among different snapshots by means of nonconvex Schatten-p quasi-norm penalty. Furthermore, for more tractable and scalable optimization problem, we apply the unified surrogate for Schatten-p quasi-norm with two-factor matrix norms. Then, a locally convergent iterative reweighted minimization method is derived and solved efficiently via a semidefinite program using the optimization toolbox. Finally, numerical simulations are carried out in the background of unknown nonuniform noise and under the consideration of coprime array (CPA) structure. The results illustrate the superiority of the proposed method in terms of resolution, robustness against nonuniform noise, and correlations of sources, in addition to its applicability in a limited number of snapshots.
Robust Automatic Speech Recognition (ASR) is a challenging task that has been an active research ... more Robust Automatic Speech Recognition (ASR) is a challenging task that has been an active research subject for the last 20 years. And still results are very modest in the highly noisy environments. In this study, we propose a new speech parameterization method based on concatenating two wavelet packet decompositions, one decomposition using low Q-factor wavelet and another with high Q-factor wavelet, to extract speech features suitable for ASR task in noisy conditions. Experiments on TIMIT dataset for phonemes recognition show that the proposed wavelet-based features outperform MFCC in all noisy conditions.
Journal of Telecommunications and Information Technology
Speech enhancement is one of the many challenging tasks in signal processing, especially in the c... more Speech enhancement is one of the many challenging tasks in signal processing, especially in the case of nonstationary speech-like noise. In this paper a new incoherent discriminative dictionary learning algorithm is proposed to model both speech and noise, where the cost function accounts for both “source confusion” and “source distortion” errors, with a regularization term that penalizes the coherence between speech and noise sub-dictionaries. At the enhancement stage, we use sparse coding on the learnt dictionary to find an estimate for both clean speech and noise amplitude spectrum. In the final phase, the Wiener filter is used to refine the clean speech estimate. Experiments on the Noizeus dataset, using two objective speech enhancement measures: frequency-weighted segmental SNR and Perceptual Evaluation of Speech Quality (PESQ) demonstrate that the proposed algorithm outperforms other speech enhancement methods tested.
The objective of this paper is to study the acoustic word stress in Arabic. We performed an acous... more The objective of this paper is to study the acoustic word stress in Arabic. We performed an acoustic analysis based on the three prosody parameters (fundamental frequency, intensity and duration), for a corpus of 100 sentences (about 400 words) in order to detect the stressed syllables. This study is based on our previous syllabification paper, and will help in determining the stress assignment rules for our Arabic Semi-Syllable Synthesizer. We discuss stress assignments on syllables of monosyllabic, disyllabic and polysyllabic words in addition to phrases with liaison cases. The assigned stress can have one of three degrees: primary stress, secondary stress and neutralized (absence of stress). In this paper, we compared the acoustic characteristics of stressed syllables with non-stressed ones. The comparison showed that the duration cue has the most prominent effect followed by intensity. F0 is the less effective cue. The primary stress is assigned to the syllable having long vowel...
Speech enhancement is one of the challenging tasks in signal processing, especially in the case o... more Speech enhancement is one of the challenging tasks in signal processing, especially in the case of non-stationary speech-like noise. In this paper we propose a new supervised speech enhancement system that uses Fischer Discriminative Dictionary Learning (FDDL) algorithm to model both speech and noise amplitude spectrum, where the cost function accounts for both "source confusion" and "source distortion" errors. In the enhancement stage, we use sparse coding on the learnt dictionary to find an estimate for both the clean speech and noise amplitude spectrum. In the final stage, a Wiener filter is used to refine the clean speech estimate. Experiments on NOIZEUS dataset using two objective speech enhancement measures: frequency-weighted segmental SNR and Perceptual Evaluation of Speech Quality (PESQ) demonstrate that FDDL outperforms other tested dictionary learning algorithms in the presence of considerable noise (0 dB) for all studied noise types, and in the presen...
Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that c... more Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that could be words, syllables or phonemes. Phonetic segmentation is about finding the exact boundaries for the different phonemes that composes a specific speech signal. This problem is crucial for many applications, i.e. automatic speech recognition (ASR). In this paper we propose a new model-based text independent phonetic segmentation method based on wavelet packet speech parametrization features and using the sparse representation classifier (SRC). Experiments were performed on two datasets, the first is an English one derived from TIMIT corpus, while the second is an Arabic one derived from the Arabic speech corpus. Results showed that the proposed wavelet packet de composition features outperform the MFCC features in speech segmentation task, in terms of both F1-score and R-measure on both datasets. Results also indicate that the SRC gives higher hit rate than the famous k-Nearest Neigh...
Journal of Telecommunictions and Information Technology
Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that c... more Speech segmentation is the process of dividing speech signal into distinct acoustic blocks that could be words, syllables or phonemes. Phonetic segmentation is about finding the exact boundaries for the different phonemes that composes a specific speech signal. This problem is crucial for many applications, i.e. automatic speech recognition (ASR). In this paper we propose a new model-based text independent phonetic segmentation method based on wavelet packet speech parametrization features and using the sparse representation classifier (SRC). Experiments were performed on two datasets, the first is an English one derived from TIMIT corpus, while the second is an Arabic one derived from the Arabic speech corpus. Results showed that the proposed wavelet packet decomposition features outperform the MFCC features in speech segmentation task, in terms of both F1-score and R-measure on both datasets. Results also indicate that the SRC gives higher hit rate than the famous k-Nearest Neighbors (k-NN) classifier on TIMIT dataset.
Received: 18 July 2020 Accepted: 6 November 2020 In this paper, we present a hybrid optimization ... more Received: 18 July 2020 Accepted: 6 November 2020 In this paper, we present a hybrid optimization framework for gridless sparse Direction of Arrival (DoA) estimation under the consideration of heteroscedastic noise scenarios. The key idea of the proposed framework is to combine global and local minima search techniques that offer a sparser optimizer with boosted immunity to noise variation. In particular, we enforce sparsity by means of reformulating the Atomic Norm Minimization (ANM) problem through applying the nonconvex Schatten-p quasi-norm (0<p<1) relaxation. In addition, to enhance the adaptability of the relaxed ANM in more practical noise scenarios, it is combined with a covariance fitting (CF) criterion resulting in a locally convergent reweighted iterative approach. This combination forms a hybrid optimization framework and offers the advantages of both optimization approaches while balancing their drawbacks. Numerical simulations are performed taking into account the...
Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and sho... more Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and short vowels which are written as punctuation marks, above or below letters. Those short vowels are normally omitted in Arabic texts because the reader can fill them and guess the meaning based on his knowledge of the language, and the context in which the words are. However, with the widespread usage of computers in linguistics application; Arabic texts need to be supplied with short vowels in order to be analyzed. Search engines, text to speech engines, and text mining tools are just some examples of applications that need Arabic texts to be vocalized before being processed. In this paper, we present a new method to supply those vocals. The approach emphasizes on unsupervised machine methods, because public Arabic corpora are not available. Arabic rich morphology and diverse orthography present serious challenges for this approach. An algorithm is developed and a system is implemented in ...
Arabic Language Processing is gaining increasing importance all around the world. This language i... more Arabic Language Processing is gaining increasing importance all around the world. This language is spoken by nearly 300 millions in the Arab World, and is an interesting language for the 1.3 billion Muslims. Arabic is becoming also a focal point of interest in many universities all around the world. The positive aspect of this interest in Arabic -which was relatively ignored when compared with other live languages-, is the intensive work on language technologies, and an increasing amount of digital contents on the Internet. Arabic Language is one of the most promoted research axis in HIAST since its foundation in 1983 (Ghneim & Al-DAkkak, 2006). In the introduction of this paper, we mention the most relevant works, in both speech and text aspects, which can be of interest in many applications, and can be a subject of resource sharing. In the second section we detail our acoustic database of semi-syllables and in the third one we develop our morphological analyzer.
This paper presents a rule-based approach to detect Arabic Speech Act types. Our approach is base... more This paper presents a rule-based approach to detect Arabic Speech Act types. Our approach is based on the fact that Arabic language is rich in morphology and that its syntax is inherently rule–based. The importance of Arabic Speech Act type recognition comes from its wide variety of application domains, such as: speech synthesis, speech and emotion recognition and conversational analysis. In this work, a rule-based Expert System has been developed in a bootstrapping manner, to classify an utterance speech act type to one of the sixteen types (interrogation, exclamation, negation, etc.). Our system was tested on a corpus of about 1500 sentences and achieved an accuracy rate of 98.92%. The expert system was used to automatically annotate a corpus (sentence-level annotation) by assigning a tag to each sentence according to its speech act type.
Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and sho... more Arabic Language has two kinds of vowels: Long vowels which are written as normal letters; and short vowels which are written as punctuation marks, above or below letters. Those short vowels are normally omitted in Arabic texts because the reader can fill them and guess the meaning based on his knowledge of the language, and the context in which the words are. However, with the widespread usage of computers in linguistics application; Arabic texts need to be supplied with short vowels in order to be analyzed. Search engines, text to speech engines, and text mining tools are just some examples of applications that need Arabic texts to be vocalized before being processed. In this paper, we present a new method to supply those vocals. The approach emphasizes on unsupervised machine methods, because public Arabic corpora are not available. Arabic rich morphology and diverse orthography present serious challenges for this approach. An algorithm is developed and a system is implemented in Java. The techniques presented in this dissertation can be applied to similar Semitic or other languages that have the same problem.
In this paper, we reformulate the gridless direction of arrival (DoA) estimation problem in a nov... more In this paper, we reformulate the gridless direction of arrival (DoA) estimation problem in a novel reweighted covariance fitting (CF) method. The proposed method promotes joint sparsity among different snapshots by means of nonconvex Schatten-p quasi-norm penalty. Furthermore, for more tractable and scalable optimization problem, we apply the unified surrogate for Schatten-p quasi-norm with two-factor matrix norms. Then, a locally convergent iterative reweighted minimization method is derived and solved efficiently via a semidefinite program using the optimization toolbox. Finally, numerical simulations are carried out in the background of unknown nonuniform noise and under the consideration of coprime array (CPA) structure. The results illustrate the superiority of the proposed method in terms of resolution, robustness against nonuniform noise, and correlations of sources, in addition to its applicability in a limited number of snapshots.
Robust Automatic Speech Recognition (ASR) is a challenging task that has been an active research ... more Robust Automatic Speech Recognition (ASR) is a challenging task that has been an active research subject for the last 20 years. And still results are very modest in the highly noisy environments. In this study, we propose a new speech parameterization method based on concatenating two wavelet packet decompositions, one decomposition using low Q-factor wavelet and another with high Q-factor wavelet, to extract speech features suitable for ASR task in noisy conditions. Experiments on TIMIT dataset for phonemes recognition show that the proposed wavelet-based features outperform MFCC in all noisy conditions.
Journal of Telecommunications and Information Technology
Speech enhancement is one of the many challenging tasks in signal processing, especially in the c... more Speech enhancement is one of the many challenging tasks in signal processing, especially in the case of nonstationary speech-like noise. In this paper a new incoherent discriminative dictionary learning algorithm is proposed to model both speech and noise, where the cost function accounts for both “source confusion” and “source distortion” errors, with a regularization term that penalizes the coherence between speech and noise sub-dictionaries. At the enhancement stage, we use sparse coding on the learnt dictionary to find an estimate for both clean speech and noise amplitude spectrum. In the final phase, the Wiener filter is used to refine the clean speech estimate. Experiments on the Noizeus dataset, using two objective speech enhancement measures: frequency-weighted segmental SNR and Perceptual Evaluation of Speech Quality (PESQ) demonstrate that the proposed algorithm outperforms other speech enhancement methods tested.
Uploads
Papers by oumayma dakkak