Categorical production in language processing helps speakers to produce phonemic contrasts. This categorization and production is utilized for the production-based and imitation-based approach in the present study. Contrastive signals in speakers' speech reflect the shapes of boundaries with categorical characteristics. Signals that provide information about lexical pitch accent contrasts can introduce categorical distinctions for productive and cognitive selection. This experiment was conducted with nine North Kyungsang speakers for a production task and nine North Kyungsang speakers for an imitation task. The first finding of the present study is the rigidity of categorical production, which controls the boundaries of lexical pitch accent contrasts. The categorization of North Kyungsang speakers' production allows them to classify minimal pitch accent contrasts. The categorical production in imitation appeared in two clusters, representing two meaningful contrasts. The second finding of the present study is that there are individual differences in speakers' production and imitation responses. The distinctive performances of individual speakers showed a variety of curves. For the HL-LH patterns, the categorical production tended to be highly distinctive as compared to the other pitch accent patterns (HH-HL and HH-LH), showing that there are more continuous curves than categorical curves. Finally, the present study shows that, for North Kyungsang speakers, imitative production is the core type of categorical production for determining the existence of the lexical pitch accent system. However, several questions remain for defining that categorical production, which leads to ideas for future research.

This study compared the relative contributions of intensity, F0, duration and vowel spectra of L2 English lexical stress by Japanese and Korean learners of English. Recordings of Japanese, Korean and native English speakers reading eighteen 2 to 4 syllable words in a carrier sentence were analyzed using multiple regression to investigate the influence of each acoustic correlate in determining whether a vowel was stressed. The relative contribution of each correlate was calculated by converting the coefficients to percentages. The Japanese learner group showed phonological transfer of L1 phonology to L2 lexical prosody and relied mostly on F0 and duration in manifesting L2 English stress. This is consistent with the results of the previous studies. However, advanced Japanese speakers in the group showed less reliance on F0, and more use of intensity, which is another parameter used in native English stress accents. On the other hand, there was little influence of F0 on L2 English stress by the Korean learners, probably due to the transfer of the Korean intonation pattern to L2 English prosody. Hence, this study shows that L1 transfer happens at the prosodic level for Japanese learners of English and at the intonational level for Korean learners.

Stressed vowels in English are usually produced using longer duration, higher pitch, and greater intensity than unstressed vowels. However, many English as a foreign language (EFL) learners have difficulty producing English lexical stress because their mother tongues do not have such features. In order to investigate if certain non-native English speakers (Korean, Japanese, and Taiwanese-Chinese native speakers) are able to produce English lexical stress in a native-like manner, speech samples were extracted from the L2 learners' corpus known as AESOP (the Asian English Speech cOrpus Project). Sixteen disyllabic words were analyzed in terms of the ratio of duration, pitch, and intensity. The results demonstrate that non-native English speakers are able to produce English stress in a similar way to native English speakers, and all speakers (both native and non-native) show a tendency to use duration as the strongest cue in producing stress. The results also show that the duration ratio of native English speakers was significantly higher than that of non-native speakers, indicating that native speakers produce a bigger difference in duration between stressed and unstressed vowels.

In this study, we examined how the upper and lower lips articulate to produce labial /p/. Using electromagnetic midsagittal articulography, we collected flesh-point tracking movement data from eight native speakers of Seoul Korean (five females and three males). Individual articulatory movements in /p/ were examined in terms of minimum vertical upper lip position, maximum vertical lower lip position, and corresponding vertical upper lip position aligned with maximum vertical lower lip position. Using linear mixed-effect models, we tested two factors (word boundary [across-word vs. within-word] and speech rate [comfortable vs. fast]) and their interaction, considering subjects as random effects. The results are summarized as follows. First, maximum lower lip position varied with different word boundaries and speech rates, but no interaction was detected. In particular, maximum lower lip position was lower (e.g., less constricted or more reduced) in fast rate condition and across-word boundary condition. Second, minimum lower lip position, as well as lower lip position, measured at the time of maximum lower lip position only varied with different word boundaries, showing that they were consistently lower in across-word condition. We provide further empirical evidence of lower lip movement sensitive to both different word boundaries (e.g., linguistic factor) and speech rates (e.g., paralinguistic factor); this supports the traditional idea that the lower lip is an actively moving articulator. The sensitivity of upper lip movement is also observed with different word boundaries; this counters the traditional idea that the upper lip is the target area, which presupposes immobility. Taken together, the lip aperture gesture is a good indicator that takes into account upper and lower lip vertical movements, compared to the traditional approach that distinguishes a movable articulator from target place. Respective of different speech rates, the results of the present study patterned with cross-linguistic lenition-related allophonic variation, which is known to be more sensitive to fast rate.

Although deep neural network-based acoustic models have greatly improved the performance of automatic speech recognition (ASR), reverberation still degrades the performance of distant speech recognition in indoor environments. In this paper, we adopt the DenseNet, which has shown great performance results in image classification tasks, to improve the performance of reverberant speech recognition. The DenseNet enables the deep convolutional neural network (CNN) to be effectively trained by concatenating feature maps in each convolutional layer. In addition, we extend the concept of multi-resolution CNN to multi-resolution DenseNet for robust speech recognition in reverberant environments. We evaluate the performance of reverberant speech recognition on the single-channel ASR task in reverberant voice enhancement and recognition benchmark (REVERB) challenge 2014. According to the experimental results, the DenseNet-based acoustic models show better performance than do the conventional CNN-based ones, and the multi-resolution DenseNet provides additional performance improvement.

A typical statistical parametric speech synthesis (text-to-speech, TTS) system consists of separate modules, such as a text analysis module, an acoustic modeling module, and a speech synthesis module. This causes two problems: 1) expert knowledge of each module is required, and 2) errors generated in each module accumulate passing through each module. An end-to-end TTS system could avoid such problems by synthesizing voice signals directly from an input string. In this study, we implemented an end-to-end Korean TTS system using Google's Tacotron, which is an end-to-end TTS system based on a sequence-to-sequence model with attention mechanism. We used 4392 utterances spoken by a Korean female speaker, an amount that corresponds to 37% of the dataset Google used for training Tacotron. Our system obtained mean opinion score (MOS) 2.98 and degradation mean opinion score (DMOS) 3.25. We will discuss the factors which affected training of the system. Experiments demonstrate that the post-processing network needs to be designed considering output language and input characters and that according to the amount of training data, the maximum value of n for n-grams modeled by the encoder should be small enough.

This paper aims to describe a knowledge-based pronunciation generation system for French. It has been reported that a rule-based pronunciation generation system outperforms most of the data-driven ones for French; however, only a few related studies are available due to existing language barriers. We provide basic information about the French language from the point of view of the relationship between orthography and pronunciation, and then describe our knowledge-based pronunciation generation system, which consists of morphological analysis, Part-of-Speech (POS) tagging, grapheme-to-phoneme generation, and phone-to-phone generation. The evaluation results show that the word error rate of POS tagging, based on a sample of 1,000 sentences, is 10.70% and that of phoneme generation, using 130,883 entries, is 2.70%. This study is expected to contribute to the development and evaluation of speech synthesis or speech recognition systems for French.

The purpose of the present study was to explore the radiological and acoustic characteristics of "Arae-a" (/${\cdot}$/) articulation in two male Jeju language speakers, focusing on selected measures in radiological images derived from computed tomography scans, as well as the first and the second formant measures in selected vowels. An elderly male speaker (a 78-year-old) and a young male speaker (a 34-year-old) participated in the study. During the production of four selected vowels, the shape of the vocal tract was identified, and selected measures were obtained from the elderly participant's computed tomography (CT) scans. For acoustic analysis, the participants were given a list of near-minimal pairs consisting of 112 words and asked to read them aloud. The results indicated that the "Arae-a" (/${\cdot}$/) articulation of the elderly speaker showed unique acoustic and radiological characteristics compared to other similar vowels, thus presenting substantial consistency with the descriptions of the "Hunminjeongeum Haeryebon." In contrast, the F1 and F2 measures of the young male's /${\cdot}$/ articulation were not distinguished from those of /ㅗ/. Current results, in part, support the scientific principles underlying the invention of "Arae-a," which reflects the shape of the vocal tract during production, and the necessity for further research.

The purpose of this study is to investigate age-specific features of 2;1- to 4;1-year -olds' production of monophthongs and diphthongs through both auditory perceptual analysis and acoustic analysis. Test material included {vowel+'da'} consisting of 7 monophthongs and 10 diphthongs and meaningful words beginning with vowels. The percentage of correct vowels was used for perceptual analysis and Praat(5.2.12) was used for acoustic analysis, analyzing variables related to monophthongs and diphthongs. The results of this study are as follows: First, perceptual analysis showed that children from an age group of 2;1 to 2;8 years showed significant difference in the accuracy level of both monophthongs and diphthongs as compared to those aged 2;9 to 3;4 years and those aged 3;5 to 4;1 years. Second, the results of acoustic analysis provided that formant (F1 and F2) of monophthong, in general, tended to decrease as age increased. In terms of F2 differentiation slope and regression slope, which were diphthong-related variables, the age group of 3;5 to 4;1 years showed a large general slope change.

Diplophonia is generally defined as the perception of more than one fundamental frequency component in a voice. Its perceptual aspect has traditionally been used to evaluate diplophonia because the perceptions can be easily evaluated, but there are limitations in the validity of the reliability of the intra- and inter-raters, examination situation, and variation of voice sample. Therefore, the purpose of this study is to confirm the reliability and accuracy of auditory perceptual evaluation by comparing non-invasive indirect assessment methods (sound waveform and EGG analysis), and to identify their usefulness with diplophonia. A total of 28 diplophonic voices and 39 non-periodic voices were assessed. Three raters assessed the diplophonia by performing an auditory perception evaluation and identifying the quasi-periodic perturbations of the acoustic waveform and EGG. Among the three discrimination methods, intra- and inter-rater reliability, sensitivity, specificity, accuracy, positive likelihood ratio, and negative likelihood ratio were examined, and the McNemar test was performed to compare the discriminant agreement. The accuracy of the auditory perceptual evaluation (86.57%) was not significantly different from that of sound waveform acoustic (88.06%), but it was significantly different from that of EGG (83.33%). The reading time (6.02 s) for the auditory perceptual evaluation was significantly different from that for sound waveform analysis (30.15 s) and EGG analysis (16.41 s). In the discrimination of diplophonia, auditory perceptual evaluation has sufficient reliability and accuracy as compared to sound waveform and EGG. Since immediate feedback is possible, auditory perceptual evaluation is more convenient. Therefore, it can continue to be used as a tool to discriminate diplophonia in clinical practice.