Search

English has three coda nasals–bilabial /m/, alveolar/n/ and velar /$\eta$/. In Chinese languages, Southern Min also has these three coda nasals and Mandarin has two-/n/and/$\eta$/, while Wu has only one coda nasal-/n/ but allows allophones in terms of pre-nasal vowels. This study was designed to explore the correlation between the production of vowel nasalization and the perception of postvocalic nasals in L2 English by Chinese learners respectively with Ll in Southern Min, Mandarin and Wu. Statistical results of the perception and the production experiments found that there was no significant correlation between the identification accuracy of postvocalic nasals and the nativelikeness of vowel nasalization in English by Chinese learners with these three Lls. Further statistical analyses revealed that coda nasal type significantly predicted the identification of nasal place of articulation in the perception experiment but not vowel nasalization in the production experiment.

Place perception for fricatives is reported to be determined by the spectral characteristics of the frication segment and also by the formant transitions in case of non-sibilants. A study is conducted to investigate the relative importance of the transitions and the frication duration on the perception of the unvoiced fricatives /f, s, sh/. Listening tests for consonant identification were conducted using vowel-fricative-vowel sequences, generated using the vowel and transition segments extracted from natural utterances and the frication segments synthesized with duration of 30–300 ms. The test results showed that the sequences with frication duration of 50 ms and larger were identified as fricatives, with the place determined by a combination of the transition and frication segments. The place perception for sequences with transition corresponding to /f/ was not affected by the frication segments, while that for sequences with transitions corresponding to /s/ and /sh/ was dominated by the frication segments. The findings indicate that estimation of the place of articulation of fricatives may be improved by combining the information from the transition and frication segments.

The Quranic alphabets can be characterized from their point of articulation (Makhaarij) and their characteristics (Sifaat). The fact is that there is limited research in the field of Tajweed from the engineering perspective. This research focused on the identification and analysis of the correct pronunciation of the Quranic alphabets on the features obtained from adult and children experts. The combination of four formant frequencies (F1, F2, F3, F4) and three Power Spectral Density values (PSD1, PSD2, PSD3) were used as the features to represent the pronunciation of each alphabets for both categories. The validation was done through a Graphical User Interface (GUI) to record the new audio signal, processes and produces the outcomes of the pronunciation in the form of percentage (%) of accuracy, as compared to the reference model of both experts' categories.

Reciting al-Quran in the correct way is an obligatory duty for Muslims, and therefore learning al-Quran is a continuous education until the correct recitation is achieved. It is important to learn Tajweed rules to master the recitation of Quranic verses. Moreover, mastering the pronunciation of Arabic sounds is the first and key step to achieve accurate recitation of al-Quran. The rules were guided by the Islamic Scholars in fields related to al-Quran from their knowledge and experiences. Very limited researches were found in the perspective of sciences and engineering. In this paper two Quranic letters (^^d0 and ^^d2) that are articulated from adjacent points of articulation were analyzed using Mel- frequency coefficient analysis. MFCCs matrices were calculated then compared using the dynamic time warping DTW technique to calculate the similarity matrices and find the similarity distance. Results show that letters from the same point of articulation have less similarity distance compared to the letters from different point of articulation.

The automatic assessment of speech plays an important role in Computer Aided Pronunciation Learning systems. However, modeling both the correct and incorrect pronunciation of each phoneme to achieve accurate pronunciation verification is unfeasible due to the lack of sufficient mispronounced samples in training datasets. In this paper, we propose a novel approach that handles this unbalanced data distribution by building multiple one-class SVMs to evaluate each phoneme as correct or incorrect. We model the correct pronunciation of each individual phoneme with a one-class SVM trained using a set of speech attributes features, namely the manner and place of articulation. These features are extracted from a bank of pre-trained DNN speech attributes classifiers. The one-class SVM model measures the similarity between the new data and the training set and then classifies it as normal (correct) or an anomaly (incorrect). We evaluated the system using native speech corpus and disordered speech corpus and compared it with the conventional Goodness of Pronunciation (GOP) algorithm. The results show that our approach reduces the false-acceptance and false-rejection rates by around 26% and 39% respectively.

Communication is a bridge between people which enables them to share facts, ideas, feelings etc. with each other. Speech communication is easier and simpler compared to other types of communication. But speech disorders affect one's ability to communicate. Dysarthria is a neuro-motor disorder, where one losses his/her ability to articulate words normally due to tongue/muscle weakness or stroke etc. It results in distorted speech which is hard to understand. Dysarthria affect more on the articulation of consonants (stops in particular) than on the articulation of vowels. Intelligibility varies greatly depending on the extent of neurological damage. As consonants are important than vowels in measuring the intelligibility, distortions in consonant production of dysarthric speech are studied and modifications are performed. This work aims at correcting devoicing of voiced stop (voiced stops are pronounced as unvoiced stops with the same place of articulation) by detecting important time instants such as glottal onset, glottal offset and burst onset thereby modifying the distortions in dysarthric speech using TMS320C6713 DSK with CC Studio 5.5. After the dysarthric speech has been recorded on the DSK 6713 processor modification is done to improve its intelligibility. Two databases were used in the experiment such as Universal Access (UA) database and Torgo database. Out of 65 bursts expected from UA database and 128 bursts expected from Torgo database, 84.6% and 82.81% bursts were detected correctly within 0.3 s intervals respectively.

Compared with corrective feedback related with pronunciation score, articulatory feedback (e.g., manner and place of articulation) is more informative and instructive in learning second language (L2). Hence, our previous work directly model and detect articulatory-level mispronunciation patterns to provide readable pronunciation erroneous tendencies (PETs) feedback. However, articulatory-level labeling is much more challenging than phone-level labelling, especially facing nonnative mispronunciations, which is pronounced by using irregular articulation manner and place. Therefore, in this paper, we proposed to use log posterior of articulatory features derived from trained DNN systems to provide labeling candidates to assist annotators make final decision. Experimental results show that the consistency rate of manual labels in our system increased from 80.7% to 92.48% In addition, the time cost for annotating each sentence reduced from about 10 minutes to 3 minutes. The reported result demonstrated the efficiency of our proposed method.

Dysarthria is a traumatic neuromotor disorder that affects the physical production of speech. It reduces the function of primary articulators that are involved in speech. Recent research has presented that the use of articulatory data provide a better assessment towards identifying the speech intelligibility of dysarthric speakers over acoustic models that are modelled based on the biological perception of speech through Mel scale than directly involving the anatomy of speech production. Articulatory data of speech, include positional data of the articulators obtained from 12 Electromagnetic Articulograph (EMA) sensor channels with each sensor channel containing 6 parameters namely x, y, z, Φ, θ, and rms. While identifying the phone intelligibility deficit for the dysarthric speakers, articulatory data of all 12 sensor combinations each with 6 parameters may create turbulence in the intelligibility assessment. Hence, an appropriate mapping of the sound units with the sensor combinations is vital. The current work, focuses on mapping the phones with the EMA sensor channels based on their place of articulation. The parameters are also grouped based on the 3D-spherical co-ordinate distribution of the articulator as (x, y, θ) and (y, z, Φ). The mapping of the EMA sensor channels with the sound units are further validated through HMM-based acoustic only models and FDA scores. The phones are trained through their corresponding articulatory sensor combination information using a 5-class support vector machine and tested through a 10-fold cross validation technique for both the parameter groups. The 5 classes include mild, moderate, moderate-to-severe, severe and normal. Each phone can be classified to any of the 5 class in both the parameter groups based on their speech intelligibility. The results show that the EMA sensor combinations with (y, z, Φ) parameter group for each phone coincide well with the acoustic and FDA scores.

The aim of this study is to compare the aspects, particularly the occurrence of vocal folds vibration during the stop closure, of the stop consonants [b], [d], and [g] produced by the Native Speakers of English and Korean EFL Speakers. It will be examined whether stop voicing in the onset and coda positions is influenced by the place of articulation. Based on K-SEC (Korean-Spoken English Corpus), i) Korean Speakers' productions of the isolated words which have the voiced stops [b], [d], and [g] as onsets, followed by six different vowels [i], [e], [s], [a], [o], and [u], and ii) the same voiced codas preceded by the aforementioned vowels are to be used for the analysis. Aspects of the initial and final stop voicing manifested by Native Speakers are also to be analyzed and then compared with those by the Korean learners of English.

In this paper the place and manner of articulation based phonological features are detected and classified. Deep Neural Network based model has been used for detection and classification task. The deep structured model is pre-trained by stacked denoising autoencoder. The system obtained 89.17% overall accuracy in detection task. In case of classification task, 50.2% of classification accuracy is observed for classifying the place of articulation based features. The manner of articulation is divided into 15 groups based on some manner based knowledge combination and classification task is performed to achieve 98.9% of classification accuracy.

For effective pronunciation error detection for second language learners, we address articulatory models based on deep neural network (DNN). Articulatory attributes are defined for manner and place of articulation. In order to efficiently train these models of non-native speech without using such data, which is difficult to collect in a large scale, we propose a multi-lingual learning method, in which the speech database of the target language (L2) and the native language (L1) of the learners are combined. We also investigate multi-task learning methods. These methods are applied to Mandarin Chinese pronunciation learning by Japanese native speakers. Effects of the multi-lingual and multi-task learning methods are demonstrated in the attribute classification of native speech and pronunciation error detection for non-native speech.

Aiming at detecting pronunciation errors produced by second language learners and providing corrective feedbacks related with articulation, we address effective articulatory models based on deep neural network (DNN). Articulatory attributes are defined for manner and place of articulation. In order to efficiently train these models of non-native speech without using such data, which is difficult to collect in a large scale, we propose a multi-lingual learning method, in which the speech database of the target language (L2) and the native language (L1) of the learners are combined. We also investigate multi-task learning methods by tuning the weights of the secondary task. These methods are applied to Mandarin Chinese pronunciation learning by Japanese native speakers. Effects of the multi-lingual and multi-task learning methods are confirmed in the attribute classification and pronunciation error detection.

The universal speech attributes to speaker verification (SV) is addressed in this paper. The manner and place of articulation form the universal attribute unit inventory, and deep neural network (DNN) is used as acoustic model. Two methods to generate the attribute units are proposed in this paper: one is that the manner and place of articulation are directly combined to generate more robust universal speech attribute units, and the other is that the different context-dependent speech arttribute units are merged to a new speech attribute unit set by means of automatic clustering in accordance with likelihood calculation. Evaluated on the core test from the 2008 NIST speaker verification evaluation (SRE), the novel generated attribute units based system can achieve a better performance than that of the phoneme based system. Furthermore, the attribute based system has demonstrated a good complementarity with the GMM-UBM/i-vector and phoneme based DNN/i-vector systems.

We propose the use of speech attributes, such as voicing and aspiration, to address two key research issues in computer assisted pronunciation training (CAPT) for L2 learners, namely detecting mispronunciation and providing diagnostic feedback. To improve the performance we focus on mispronunciations occurred at the segmental and sub-segmental levels. In this study, speech attributes scores are first used to measure the pronunciation quality at a sub-segmental level, such as manner and place of articulation. These speech attribute scores are integrated by neural network classifiers to generate segmental pronunciation scores. Compared with the conventional phone-based GOP (Goodness of Pronunciation) system we implement with our dataset, the proposed framework reduces the equal error rate by 8.78% relative. Moreover, it attains comparable results to phone-based classifier approach to mispronunciation detection while providing comprehensive feedback, including segmental and sub-segmental diagnostic information, to help L2 learners.

L2 learners of Mandarin have difficulty learning native-like pronunciation of nasal codas. In order to help them learn native-like pronunciation, we propose to develop targeted classifiers for automatic pronunciation error detection. In this paper, perceptual experiments with modified speech are designed to analyze the exact position of the landmark of a nasal coda. Based on perceptual results from isolated words, we propose that information about nasal coda place of articulation is most dense near a landmark at the center of the nasalized vowel. Landmarks detected in a database of Japanese learners of Mandarin, and classified as correct vs. incorrect using an SVM. The result shows that the detection performance of the SVM+Landmark system is similar to that of a DNN-HMM+MFCC system. When the two systems are combined, an FRR of 4.6% is achieved at DA of 83.9%. This performance is comparable to that of previously developed classifiers for 16 common Mandarin pronunciation errors.

Place of articulation obtained by analysis of the speech signal is useful for visual feedback of articulatory efforts for speech training of hearing impaired children and for improving pronunciation by learners of second languages. Its estimation by direct imaging of the oral cavity is needed for validating the estimation from the speech signal. For such applications, an automated technique is presented for estimating the place of articulation by graphical processing of the upper and lower contours of the oral cavity image. It iteratively estimates the axial curve as an axis of symmetry of the oral cavity, such that the curve approximately bisects the normals to it. Distance between the contours along the normal to the axial curve gives the oral cavity opening and position of the smallest opening provides the place of articulation. The values estimated using the automated technique closely matched those obtained by manual marking of the visually estimated place of maximum constriction for the oral cavity images of vowels, stops, and fricatives, from the XRMB and MRI databases.

This paper investigates the perceptual confusion pattern in the three coda nasals /-m, -n, -η/ in Southern Min for insight into the effect of tone on the identification of nasal place of articulation. A general hierarchy of perceptual confusability - m>-η>-n was found in the current set of data, which is compatible with the fact that /-m/ was lost first in the historical merger of coda nasals in Chinese. Tones with a high pitch at the ending point resulted in higher accuracy of nasal identification in complete syllables. However, there was no tonal effect on the identification of nasal murmur. Vowel duration and vowel-nasal formant transition were the perceptual cues that listeners mostly relied on.

The aim of the study is to experimentally verify the characteristics of Assamese consonants. All the previous work has been done subjectively. This study is the first attempt to objectively verify the characteristics of Assamese consonants. There are eight vowels, two semivowels and twenty one consonants in standard colloquial Assamese. Semivowels are excluded in this study. The study of place and manner of articulation is based on the Electropalatography system and acoustic data of the phoneme. Phonation process is determined by Electroglottograph. The study reveals that in Assamese there are contrasts in three distinct places of articulation: the lips, the alveolar ridge and the velum. It is also observed that alveolar plosives in context with vowel /u/ and /a/ show signature of retroflection.

This paper addresses the problem of Query-by-Example Spoken Term Detection (QbE-STD). Posteriorgrams have been widely used in the research on QbE-STD. Features based on articulatory classes are known to be robust to phonemic variations. The articulatory features like voicing and place of articulation are the main distinguishing features among some plosives and fricatives. These properties inspire the study of posteriorgrams based on articulatory classes for QbE-STD. Most of the previous works based on articulatory features have defined a large number of articulatory classes making it difficult to use them directly for pattern matching. Also, most of the works have completely ignored the uniqueness of the phonemes having transitory places of articulation eg. diphthongs and approximants. These issues have been addressed in this work while carefully selecting low dimensional articulatory motivated (LDAM) posteriorgrams on the basis of detailed experiments. This work is the first to show that the articulatory based posteriorgrams can outperform the phonemic posteriorgram significantly in a stand alone way (without any support from acoustic or phonemic features) for the task of QbE-STD.

Hidden Markov Model (HMM) has contributed greatly in the area of speech recognition during last two decades. However, in recent days, detection-based, bottom-up speech recognition techniques achieve high success rate. In this detection-based, bottom-up approach of speech recognition, first step is detection of speech attributes like place and manner of articulation of the phonemes. This paper describes about the detection of attributes which leads to identification of place and manner of articulation of Bengali phonemes using Deep Neural Network (DNN).

Marathi is a language spoken predominantly in the state of Maharashtra, India and has close to 70 million native speakers. It is based on the Devanagiri script, which the script used for the mother of all Indic languages, Sanskrit. Devanagiri consists of 46 phonemes and arranges them in a scientific manner based on the process of generation. The consonants are arranged in varnas (classes), and each varna has a unique place of articulation. This work focuses on identifying individual class variants of the first five varnas of the Marathi Devanagiri script.

This paper attempts to provide some insights about the relationship between the differentiability and the classification importance of consonants in Chinese speech communication. The two characteristics can be modelled by the perceptual distance and the functional load respectively. We have a clustering analysis of Chinese consonants based on functional load (FL) relied on mutual information (MI) between the text and its phoneme transcription. Then we compare our clustering result with that based on the perceptual distance by articulation tests. By experimenting on the Chinese newspaper corpus with millions of sentences, we find most phonemes at the same place of articulation with different manners tend to have large FLs pairwise. It is consistent with the result that those phonemes tend to have long perceptual distance pairwise.

In spite of the great success of the deep neural network (DNN) in speech processing, it is still unclear what kind of underlying mechanisms are involved in this achievement. This preliminary study attempts to find an answer by investigating the functions of DNN's hidden layers in representing speech articulations. Two sets of experiments are performed on the hidden layers in speech recognition. The layer removing experiment is conducted on the English TIMIT database, and the layer replacing experiment is to substitute a layer in an English DNN by the corresponding layer in a Japanese DNN. It is found that the different layers seem to be responsible for different phoneme groups according to the place of articulation. The lower layers are responsible for the back vowels, and the higher layers are responsible for the front vowels. The second layer (i.e. the first hidden layer) of the seven-layer network has major responsibility for more than half of the consonants with the constriction located in the front of the vocal tract, while the other consonants rely on the middle and higher layers. The layer replacing experiment demonstrated that the above relation was language independent. It is necessary to design elaborate studies to discover more details in the future.

In computer assisted pronunciation training system (CAPT), detecting mispronunciations produced by non-native speakers and providing detailed instructive feedbacks are two key points, because it is helpful to L2 learners to improve their pronunciation more effectively. We wish to give feedbacks relate to articulation-placement and articulation-manner, and connect with the detection task through modeling and detecting the mispronunciations on account of the erroneous articulation tendencies, including the place of articulation and the manner of articulation. In this paper, we focus on Japanese learners of Mandarin Chinese, and present our study on phonetic segmental mispronunciations detection work. The experimental results show that the approach can detect the mostly representative pronunciation errors moderately well.

A finite set of phonetic units is used in human speech, but how our brain recognizes these units from speech streams is still largely unknown. The revealing of this neural mechanism may lead to the development of new types of speech brain computer interfaces (BCI) and computer speech recognition systems. In this study, we used electrocorticography (ECoG) signal from human cortex to decode phonetic units during the perception of continuous speech. By exploring the wavelet time-frequency features, we identified ECoG electrodes that have selective response to specific Chinese phonemes. Gamma and high-gamma power of these electrodes were further combined to separate sets of phonemes into clusters. The clustered organization largely coincided with phonological categories defined by the place of articulation and manner of articulation. These findings were incorporated into a decoding framework of Chinese phonemes clusters. Using support vector machine (SVM) classifier, we achieved consistent accuracies higher than chance level across five patients discriminating specific phonetic clusters, which suggests a promising direction of implementing a speech BCI.

In the last few decades, articulatory trajectories have been used to improve the accuracy of speech recognition and synthesis. In this article, we directly adopt the trajectories to control coarticulation in plosive synthesis to demonstrate that our 2-D Digital Waveguide Mesh (DWM) is able to simulate formant transition in plosives accurately. The formant transitions in our results are the acoustic results of controlling four places of articulation trajectories using electromagnetic (midsaggital) articulograph (EMA) data. Positions of lip, tongue tip, tongue body and tongue dorsum are inversely proportional mapped to cross-sectional area function of those places. Linear interpolation between them enabled all tract movement to be modelled. The results show that tract movements can be modeled in a linear or non-linear fashion and the results demonstrate that non-linear coarticulation provides the most appropriate outputs for plosives.

In this analysis paper, we investigate the effect of phonetic clustering based on place and manner of articulation for the enhancement of throat-microphone speech through spectral envelope mapping. Place of articulation (PoA) and manner of articulation (MoA) dependent GMM-based spectral envelope mapping schemes have been investigated using the reflection coefficient representation of the linear prediction model. Reflection coefficients are expected to localize mapping performance within the concatenation of lossless tubes model of the vocal tract. In experimental studies, we evaluate spectral mapping performance within clusters of the PoA and MoA using the log-spectral distortion (LSD) and as function of reflection coefficient mapping using the mean-square error distance. Our findings indicate that highest degradations after the spectral mapping occur with stops and liquids of the MoA, and velar and alveolar classes of the PoA. The MoA classification attains higher improvements than the PoA classification.

This paper studies on the Stops CV Structure in Amdo Tibetan's Xiahe Dialect. This study uses the research methods of speech acoustics, extraction of speech formant parameters. To research the CV structure Intra-Syllable Coarticulation. The main results of the research were as follows: In Xiahe Dialect CV structure exist coarticulation effects between consonants and vowels. The consonant made a great impact on vowel, relating mainly to the place of articulation of consonants. Consonant initials also influenced by the vowel. Through the locus slope of consonant, we found that place of articulation at back of consonants were affected most, unaspirated sound is greater than the aspirated sound.

In this study, the Vowel-to-Vowel (V-to-V) coarticulatory effect in the Vowel-Consonant-Vowel (VCV) sequences is investigated. The stimuli are in the form of C1V1.C2V2, which is designed for V1 to affect V2. The subjects are twelve native speakers of standard Chinese, and the F2 onset value, as well as the F2 delta value, is analyzed. Results show that, due to long temporal separation between vowels for the aspirated stops relative to the unaspirated ones, the effect of aspiration is obvious, and in the contexts of aspirated stops, the effect of place of articulation is restrained.

In this study the prosodic effect on the temporal realization of segments in Chinese is analyzed. The test segments are in the form of `C1V1#C2V2', with `#' varied to be intonational phrase boundary and intonational phrase medial position, and C2 varied to be labial, alveolar and velar stop. The subjects are fourteen native speakers of standard Chinese. The VOT of the consonants and the durations of the vowels are analyzed. Results show that the effect of prosody on the duration of the preceding vowel is great, and there is also an effect of place of articulation of the second consonant on the duration of the first vowel, the VOT of the second consonant and the duration of the second vowel.

Computer-assisted pronunciation training (CAPT) was introduced for language education in recent years. CAPT scores the learner's pronunciation quality and points out wrong phonemes by using speech recognition technology. However, although the learner can thus realize that his/her speech is different from the teacher's, the learner still cannot control the articulation organs to pronounce correctly. The learner cannot understand how to correct the wrong articulatory gestures precisely. We indicate these differences by visualizing a learner's wrong pronunciation movements and the correct pronunciation movements with CG animation. We propose a system for generating animated pronunciation by estimating a learner's pronunciation movements from his/her speech automatically. The proposed system maps speech to coordinate values that are needed to generate the animations by using multilayer perceptron neural networks (MLP). We use MRI data to generate smooth animated pronunciations. Additionally, we verify whether the vocal tract area and articulatory features are suitable as characteristics of pronunciation movement through experimental evaluation.

Magnetic Resonance Imaging technique makes it possible to measure the motion of tissues in our body organs more clearly than other medical imaging techniques. The aim of this paper is to build a co-articulatory model based on Magnetic Resonance Images (MRI). This work is a blend of various emerging technologies such as computer vision based visualization technologies, cognitive science, medical science, speech recognition. The sounds of human speech can be combined in many ways, and the associated articulator movements vary as the kinematic context changes. This kinematic variation, known as co-articulation is one of the most pervasive characteristics of speech production. Visualization of co-articulatory effects involved in the speech production will lead to a better understanding of the speech production process. MRI video obtained (from the subject AR) during the co-articulation of Tamil phonemes has been incorporated as input and processed to envision the movements of the key articulators involved in the speech production process. The Region of Interest for the articulators such as jaw, tongue, lower lip, and upper lip were obtained. The motion parameters for individual articulators and their positions in subsequent frames are estimated using Block matching algorithm. Estimated motion parameters are visualised and then reproduced. This system can act as an efficient tool to control the place of articulation visually to aid second language learners and also for the people suffering from mis-articulation to learn the correct method of articulation.

This paper discussed the effect of vowel (fatha, kasra and damma) in Arabic consonants. These vowels are added to the basic consonants with three simple diacritics using the utterances of every dental consonant concerned by Malaysian children. The dental consonants refer to the consonants utter using dental medium. It is called the place of articulation at labiodentals, dental and interdental. The formant frequencies produced for each place of articulation are study based on their spectrogram. The visualization of the spectrogram makes the formants easily being identified by normal human being’s vision. The formants (F1, F2 and F3) are averaged and the results show the different increment and decrement compare to utterances of single phoneme. F1 is decreased for all observed consonants. Meanwhile F2 and F3 changes according to its manner of articulation, where in this study the tabulation are plosive and fricative.

Both frication noise and formant transitions have been argued to be important cues in sibilant perception. Additionally, the availability of coarticulatory cues to adjacent segments in vowel transitions has been argued to be asymmetric; more coarticulatory information is available in vowel onsets. Less research has examined how the use of cues may vary across syllable positions. This paper investigates cue-weighting in sibilant perception by Taiwanese Mandarin listeners. Specifically, we examine whether listeners' attention is drawn to different cues when the sibilant occurs in a phonotactically illegal position. The results demonstrate that listeners are more sensitive to sibilants distinctions in phonotactically illegal contexts (VC sequences).We conclude with the observation that while the reliance on frication noise seems to be greater in phonotactically illegal contexts, in phonotactically legal contexts, place of articulation cues from the frication noise and vowel transitions are more integrated.

Computer Assisted Pronunciation Training (CAPT) systems can judge the overall pronunciation quality and point out the pronunciation errors by recognition of the user's utterance to improve the users' oral ability. However, the performance of the current error detection systems can't satisfy the users' expectation. This paper carefully designs the pronunciation dictionary of the speech recognition engine of the CAPT system based on observation of the spectral properties of some phonemes. First, the phonetic symbol /R/ utilized in some modern English dictionaries is replaced by two symbols, /R/ (as in word result) and /ER/ (as in word bear), due to the large spectral deviation in these two cases. Moreover, rather than using a consonant sequence to represent a consonant cluster, we use one symbol to represent the whole for some consonant clusters, according to the phonetic properties of the consonants. Finally, we divide the phoneme /l/ into clear /l/ and dark /l/ according to the distinction in the manner of articulation and the place of articulation. So is /n/. As a result, we get a new pronunciation dictionary which is more suitable for the speech recognition engine of the CAPT system. The HMMs are trained by using the TIMIT database, and evaluated on a database involving 40 Chinese undergraduates. Experimental results illustrate the effectiveness of the new pronunciation dictionary.

The aim of this study is to determine the place of articulation for Arabic phonemes by subjects on non-arabic spoken language namely Malay. Every phoneme must be pronounced accurately in order to obey the rule of tajweed of holy Quran. By using Fourier analysis technique, the sound waveform is transform into spectrum which is frequency representation of the signal. The spectrogram is used to determine the formant frequency. The spectrogram for each subject is observed to determine its formant frequency. This study shows that second and third formant frequency (F2 and F3) increased as the articulation is made towards into the mouth.

It is well known that bursts and voiced formant transitions serve as separate cues to the place of articulation of initial stop consonants. The Vietnamese presents three final voiceless stop consonants /p, t, k/ without bursts. It is an opportunity to study these final stop consonants and to compare their characteristics with those of the corresponding initial stop consonants. As final consonants were never studied before, this paper analyses the vowel-consonant (VC) and consonant-vowel-consonant (CVC) productions in terms of the transition duration, the starting formant transition values and the slopes of the VC transitions. Measurements have shown that in the same preceding vowel contexts, the three final stop consonants /p, t, k/ are always clearly different by at least one of the three slopes of F1, F2, and F3. These final stop consonants can also be differentiated in the locus equation space. The results also pointed out the effects of the final consonants on either long vowels or short vowels. This explains why Vietnamese could not pronounce the short vowels in isolation.

Voice onset time (VOT) has been considered as one of the crucial methods for examining the timing of voicing in stop consonants in many languages. This study was designed to examine VOT production for phonetically voiceless stops of Mandarin and English by native Chinese speakers. Thirty-six Taiwanese Chinese speakers participated in this study. The results indicates that: 1) Based on the three universal categories proposed by Lisker and Abramson (1964), for phonetically voiceless stops, Mandarin and English occupies the same place along the VOT continuum; 2) The mean VOT value for the apical stop /t/ is slightly lower than the mean value for the labial stop /p/. This does not conform to the general agreement, that is, the further back the place of articulation the longer VOT.; 3) The difference between the mean VOT values of English /p/ and /t/ produced by Chinese speakers is subtle, whereas the difference reaches significance for native English speakers. Therefore, it is suggested that first language (i.e. Chinese) could be a crucial factor affecting L2 (i.e. English) productions. Moreover, variations, both within and between speakers, could be further examined in future studies.

We propose a rescoring framework for speech recognition that incorporates acoustic phonetic knowledge sources. The scores corresponding to all knowledge sources are generated from a collection of neural network based classifiers. Rescoring is then performed by combining different knowledge scores and they are used to reorder candidate strings provided by state-of-the-art HMM-based speech recognizers. We report on continuous phone recognition experiments using the TIMIT database. Our results indicate that classifying manners and places of articulation provides additional information in rescoring, and improved accuracies over our best baseline speech recognizers are achieved using both context-independent and context-dependent phone models. The same technique can be extended to lattice rescoring and large vocabulary continuous speech recognition.

In human heads there is a strong structural linkage between vocal tract and facial behavior during speech. For a robotic talking head to have a human-like behavior, this linkage should be emulated. One way to do that is to compute an estimate of the articulatory features which produce a given utterance and then to transform them into facial animation. We present a computational model of human vocalization which is aimed at describing the articulatory mechanisms which produce spoken phonemes. It uses a set of fuzzy rules and genetic optimization. The former represents the relationships between places of articulations and speech acoustic parameters, while the latter estimates the degrees of membership of the places of articulation. That is, the places of articulation are considered as fuzzy sets whose degrees of membership are the articulatory features. The trajectories of articulatory parameters can be used to control a graphical or mechanical talking head. We verify the model presented here by generating and listening to artificial sentences. Subjective listening tests of artificially generated sentences from the articulatory description resulted in an average phonetic accuracy of about 79 %. Through the analysis of a large amount of natural speech, the algorithm can be used to learn the places of articulation of all phonemes of a given speaker

Display of vocal tract shape can be used in speech training aids for the hearing impaired children, as it provides a visual feedback of the articulatory efforts. Estimation of vocal tract shape, based on LPC and other analysis techniques, works satisfactorily for vowels but generally fails during stop closures. Indication of correct place of articulation is very important, particularly for non-labial consonants. In order to study the dynamics of the vocal tract shape estimation during transitions at vowel-consonant boundaries, we have used "areagram", a spectrogram-like two-dimensional (2D) display of estimated vocal tract cross-sectional area as a function of time and position along the tract length. Area estimation is based on reflection coefficients obtained from LPC analysis of speech. Based on estimated area during the transition segments preceding and following the stop closure, bivariate polynomial surfaces are obtained and these are used for estimation of the vocal tract shape during stop closure by performing 2D interpolation. The place of closure for various stop consonants could be estimated satisfactorily from the conic surface approximation.

Articulatory feature modeling in automatic speech recognition (ASR), while not (yet) mainstream, has received a significant amount of attention in recent research ((S. Chang, et al., 2001), (K. Kirchhoff, May 2000), (S. Stuker, et al., 2003), (F. Metze and A. Waibel, 2002) inter alia). One study in particular (S. Chang, et al., 2001) has provided evidence that hierarchical articulatory feature models can potentially significantly outperform their non-hierarchical counterparts. In such a system, the probability of an articulatory feature is conditional upon some other feature - for example, the classifier for place of articulation may depend on the manner of articulation. In this work, we seek to further the studies in (S. Chang, et al., 2001) by changing the assumption of perfect recognition of the conditioning class made in that study. The gains shown over non-hierarchical classification are minimized; our analysis shows that this is in part because the errors in different acoustic feature streams are in fact correlated. We conclude the study by observing that joint acoustic feature modeling, rather than conditional modeling, may provide better gains

Speech contains important information in the nonstationary transitions between sounds. The Wigner-Ville distribution (WVD) allows a time-frequency representation of these transients, which convey acoustic cues or information about the sound identity, place of articulation, and speech anomalies. Speech has a complex spectral structure, which requires a careful choice of WVD analysis and time smoothing windows to attenuate spectral interferences. The smoothed WVD (SWVD) analysis of synthetic and real continuous speech transitions accurately shows the formant dynamics, glottal excitation, and energy variations. A simple classification experiment based on nasal consonant to front vowel transition SWVD spectra is presented. >

The Boltzmann machine algorithm and the error back propagation algorithm were used to learn to recognize the place of articulation of vowels (front, center or back), represented by a static description of spectral lines. The error rate is shown to depend on the coding. Results are comparable or better than those obtained by us on the same data using hidden Markov models. The authors also show a fault tolerant property of the neural nets, i.e. that the error on the test set increases slowly and gradually when an increasing number of nodes fail. >

Some aspects of training methods in building a predictive classification model are studied in the context of discriminating between the stop consonants based on their places of articulation. Two training algorithms are taken-up: one from the likelihood training category and the other from the discriminative training category. These algorithms are discussed in a unified framework using information theoretic concepts to bring out their advantages and limitations with reference to the classification task. The experimental studies on recognition confirm the usefulness of exploiting these advantages for modelling the stop consonant transition region by a predictor. A waveform based scalar predictor, trained by the modified maximum likelihood (first) algorithm shows a substantial reduction in pitch pulse interference. The significance of discriminatively training the predictive classifier using a maximum mutual information type (second) algorithm is brought out in a spectral prediction framework.

The phonetic and linguistic features of spoken Chinese which are closely related to the research and development of speech I/O systems are discussed at different linguistic levels. It is shown that: (1) the syllable structures of Chinese are an advantage in increasing the syllable intelligibility; (2) the rank order of the principal phonetic features is voiced, aspirated, fricative, and places of articulation; (3) the statistical relation between the intelligibility of different speech units can be described by a Bayes probability model; (4) the task of developing a continuous speech recognition system for Chinese faces a number of linguistic barriers. >

Presents an approach to classification of place of articulation (POA) of stop consonants in continuous speech. As the clues for POA are dependent on the formant transitions between a consonant and its following vowel, the authors propose to classify the CV transition rather than just the POA. They also propose pitch region analysis based signal processing methods to extract parameters suitable for classification of CV transitions. They present the results of their studies on classification using the multilayer perceptron model and compare the performance for different parametric representations. >

This paper investigates the effectiveness of using neural networks in classifying Malay speech sounds according to their place of articulation and voicing. The system is very different from conventional speech recognition systems, where the systems do not classify the speech sounds into groups of voicing and place of articulation. The system proposed classifies 16 selected Malay syllables into their groups of phonetic features. The Malay syllables are initialized with stops and followed by succeeding vowels. The speech tokens are sampled at 16 kHz with 16-bit resolution. LPC-derived cepstrum is used to extract the speech features. A three-layer multilayer perceptron (MLP) is used to train and recognize the Malay syllables. The system gives an encouraging result, with an average accuracy of 92.92%.

The paper aims at investigating the use of Neural Networks in classifying the Malay plosives based on their places of articulation. The approach is quite different from the conventional speech recognition methods, where they classify the speech sounds according to their phonemes, syllables or isolated words. A three-layer Multi-layer Perception (MLP) is used to classify six plosive sounds in various phonetic contexts. The MLP is trained with stochastic Back-propagation (BP), where the weights are updated after presentation of each training token. The speech tokens are sampled at 16 kHz with 16-bit resolution. The proposed method can be used in a speech training system to show the tongue placement based on the speech sounds. The method is able to achieve a promising recognition rate of 90% when tested speaker-independently on adult and children speakers.