Speaker Characteristics Research Articles

Group Sparse Representation With WaveNet Vocoder Adaptation for Spectrum and Prosody Conversion

The statistical approach to voice conversion typically consists of a feature conversion module followed by a vocoder. So far, the feature conversion studies are mainly focused on the conversion of spectrum. However, speaker identity is also characterized by prosodic features, such as fundamental frequency F0 and energy contour among others. In this paper, we study the transformation of speaker characteristics both in terms of spectrum and prosody. We propose two novel techniques that effectively use a limited amount of source-target training data and leverage a large general speech corpus to improve the voice conversion quality. First, we study the phonetic sparse representation under the group sparsity mathematical formulation. We use phonetic posteriorgrams PPGs together with spectral and prosody features to form tandem feature in the phonetic dictionary. The tandem feature allow us to estimate an activation matrix that is less dependent on source speakers, thus providing a better voice conversion quality. Second, we study the use of WaveNet vocoder that can be trained on general speech corpus from multiple speakers and adapted on target speaker data to improve the vocoding quality. We benefit from the large general speech databases that are used to train the PPG generator, and the WaveNet vocoder. The experiments show that the proposed conversion framework outperforms the traditional spectrum and prosody conversion techniques in both objective and subjective evaluations.

P‐101: Analyzing Acoustic Characteristics of Multi‐channel Speaker Directly Driving Flat Panel Display: Considering the Acoustic Stereo Effects

With the technological improvements in the display industry, organic light‐emitting diode (OLED) panel manufacturers have changed screen high‐definition to simple 3D effects, augmented reality, or virtual reality. In previous research, we introduced the sound technology that vibrates the OLED panel, which name is crystal sound OLED(CSO). The difference from LCD panels, OLED panels are the simple structure, that are consists of the glass of screen and that the backside for supports. Therefore, the screen panel can take on the role of a diaphragm of an ordinary dynamic speaker. Human hearing can confirm the sound position of the three‐dimensional space by utilizing the delay difference of the sound reaching the left and right ear. If we generate these characteristics virtually, can create a 3D stereo effect. In another previous research, It is possible to create a stereo sound with multiple independent speakers. However, there is a limitation in the sound of a display speaker composed on a single panel. In this study, we analyze the acoustic characteristics of a flat speaker to realize such stereophonic sound and improve sound quality.

Quality at the American Society for Radiation Oncology Annual Meeting: Gender Balance Among Invited Speakers and Associations with Panel Success

The American Society for Radiation Oncology (ASTRO) annual meeting is the most prominent international conference in radiation oncology. It represents one of the greatest time efforts for ASTRO volunteers, and the quality of the annual meeting is a high priority. Measures of diversity have been linked with quality of academic efforts. We conducted a study of gender diversity of the ASTRO invited speakers, focusing on speaker and panel characteristics, and associations of these characteristics with measures of audience satisfaction. We created a database of all invited speakers (n = 1499 cumulative speaking positions, n = 725 individual speakers) and panels (n = 381) in the ASTRO annual meetings from 2012 to 2016. Speaker characteristics were acquired using publically available online search tools (including Scopus for citation metrics). SAS software was used for statistical analysis. Of the 725 individual speakers, 27% were women. Men had higher median Hirsch index (H-index) (27 vs 20, P < .001), M-index (1.36 vs 1.11, P < .001), earlier first publication (1994 vs 1997), professorship (57% vs 40%, P < .001), chair status (25% vs 14%, P < .001) and fellow designations/distinctions (ie, FASTRO/FACR/FAAPM) (12% vs 9.5%, P< .17). Median panel gender distribution was 25% female, and greatest female representation was in breast, gynecology, and pediatrics. On multivariable analysis, panel characteristics associated with individual measures of audience satisfaction included single discipline presentations (metric: overall evaluation score), panels not proposed or endorsed by committee (metric: room count), higher H-index (metric: room count) and M-index (metric: request for session repeat), and panels with gender composition >50% female (metric: request for session repeat). Female conveners were associated with greater panel gender diversity. Given observations that women constitute a minority of invited speakers and that greater female representation correlates with a key panel success metric, efforts to ensure greater diversity of those included in AM speaking invitations merit serious attention.

Text-independent speaker identification based on deep Gaussian correlation supervector

Great progress has been made in speaker recognition by extracting features from Gaussian mixture model (GMM) or deep neural network (DNN). In this paper, to extract the personality characteristics of speakers more accurately, we propose a novel deep Gaussian correlation supervector (DGCS) feature based on a DBN-GMM hybrid model. In the method, we firstly extract MFCC from preprocessed speech signals and employ a DBN to gain bottleneck features. Then bottleneck features are fed to a GMM to extract deep Gaussian supervector (DGS) which can be as the input of SVM achieving pattern discrimination and judgment. Further considering the relevance between deep mean vectors of DGS, DGS will be transformed to DGCS by the method of supervector recombination. Our experiments show that utilizing DGCS can significantly improve recognition rate by 17.979% compared to the system only with supervector, 18.22% compared to the system with DGS and 1.875% compared to the system with correlation supervector. In addition, the proposed DGCS demonstrates that time complexity for identification task can be largely reduced.

Accommodating linguistic prejudice? Examining English teachers’ language ideologies

PurposeThe purpose of this study is to support the integration of scientifically grounded linguistic knowledge into language teaching in English Language Arts (ELA) classrooms through building an understanding of what teachers currently know and believe about language.Design/methodology/approachIn total, 310 high school English teachers in the USA responded to a survey about their language beliefs. Statistical analysis of responses identified four distinct constructs within their belief systems. Sub-scales were created for each construct, and hierarchical regressions helped identify key characteristics that predicted beliefs along a continuum from traditional/hegemonic to linguistically informed/counter-hegemonic.FindingsKey findings include the identification of four belief constructs: beliefs about how language reveals speaker characteristics, beliefs about how society perceives language use, beliefs about how language should be treated in schools and beliefs about the English teacher’s role in addressing language use. In general, teachers expressed counter-hegemonic beliefs for their own role and their view of speaker characteristics. They expressed hegemonic beliefs for societal perceptions and the dominant school language narrative. Taking a linguistics class was associated with counter-hegemonic beliefs, and teaching longer was associated with more hegemonic beliefs.Practical implicationsThe findings of this study suggest that the longer teachers teach within a system that promotes hegemonic language practices, the more they will align their own beliefs with those practices, despite having learned linguistic facts that contradict pervasive societal beliefs about language. The Dominant School Language Narrative currently accommodates, rather that disrupting, linguistic prejudice.Originality/valueA current understanding of teachers’ language ideologies is a key step in designing teacher professional development to help align teaching practices with established linguistic knowledge and to break down a socially constructed linguistic hierarchy based on subjective, and frequently prejudicial, beliefs.

Who can speak that language? Eleven-month-old infants have language-dependent expectations regarding speaker ethnicity.

Research demonstrates that young infants attend to the indexical characteristics of speakers, including age, gender, and ethnicity, and that the relationship between language and ethnicity is intuitive among older children. However, little research has examined whether infants, within the first year, are sensitive to the co-occurrences of ethnicity and language. In this paper, we demonstrate that by 11 months of age, infants hold language-dependent expectations regarding speaker ethnicity. Specifically, 11-month-old English-learning Caucasian infants looked more to Asian versus Caucasian faces when hearing Cantonese versus English (Studies 1 and 3), but did not look more to Asian versus Caucasian faces when paired with Spanish (Study 2), making it unlikely that they held a general expectation that unfamiliar languages pair with unfamiliar faces. Moreover, infants who had regular exposure to one or more significant non-Caucasian individuals showed this pattern more strongly (Study 3). Given that infants tested were raised in a multilingual metropolitan area-which includes a Caucasian population speaking many languages, but seldom Cantonese, as well as a sizeable Asian population speaking both Cantonese and English-these results are most parsimoniously explained by infants having learned specific language-ethnicity associations based on those individuals they encountered in their environment.

Acoustic radiation characteristics improvement according to the shape change of flat-plate display exciter speaker

For human, sound and video are evolving as the information transmission method. People are quick and easy to understand when sound and video are transmitted at the same time. In previous studies, we introduced a study to increase the quality of sound by equalizing the position of sound and sound in a flat panel display such as OLED (organic light emitting diode) TV. In that, we have implemented a sound field by separating two or more channels by vibrating the display panel of an OLED display. In this study, we investigated the effect of the variation of the shape of the pole piece on the radiation characteristics of the exciter speaker. In dynamic speakers, a pole piece guides the voice coil. However, in the exciter, the Pole Pieces excite the diaphragm directly. Initially, due to the limitations of exciter manufacturing technology, two exciters were placed separately. The first improvement was arranged on the same axis, and the sound was implemented by twin structure. However, because of the use of two pole pieces, the subtle phase difference of the two influenced the radiation pattern. In this study, the elliptical pole piece was introduced to improve the radiation characteristics in the transverse axis direction. The use of a single pole piece not only improves acoustical characteristics in the longer radius direction; furthermore, it is confirmed that the sound quality is improved by reducing manufacturing and driving errors even in the radial direction.

Code Mixing and Code Switching in Contemporary Indonesia Dangdut Music: A Sociolinguistic Case of Nela Kharisma Song Lyrics

Language was basically a medium for describing ideas as well as when singers singing their songs by using language. But listening to Nella Kharisma’s song lyrics were very interesting because the songs used many phrases that were translated into other languages and used code mixing in the lyrics of the songs. Language, in the view of sociolinguistics contained different variations. A sociolinguist explained the correlation between variations with social and situational factors. This situation affected the language used by musicians whom at any time incorporated linguistic elements outside of their native language. Due to the variety of language usage, what had come to be called bilingual users has emerged. Sociolinguistics as a branch of linguistics viewed or placed language position in relation to language using. The tendency in switching between languages using due to the changing of situations was called code switching. The discussion of code switching was usually followed by code mixing, which was about the existence of a basic code that was used and had autonomous functions. A speaker who had mastered more than one language would have the opportunity to mix codes. But that did not mean that speakers who mastered more languages always mix more codes. A speaker who slipped a lot of other languages into the language he was using was called code mixing. The background of code mixing and code switching could be seen from the mixed language of a person/figure using more than one language. Specific characteristics of speakers would influence the code mixing. This discussion became very interesting when looking at the lyrics of the songs sung by Nella Kharisma and looking at them from sociolinguistic parameters. This paper would talk about the problem of switching from one language to another, using sociolinguistic approaches method and linguistic analysis method. The results achieved were to find out how far the language was used by Nella Kharisma in the lyrics of the song she sang. Keywords : song lyrics, code switching, code mixing, Indonesia dangdut DOI : 10.7176/JLLL/54-09 Publication date :March 31 st 2019

Age Estimation in Foreign-accented Speech by Native and Non-native Speakers.

Current research shows that listeners are generally accurate at estimating speakers' age from their speech. This study investigates the effect of speaker first language and the role played by such speaker characteristics as fundamental frequency and speech rate. In this study English and Japanese first language speakers listened to English- and Japanese-accented English speech and estimated the speaker's age. We find the highest correlation between real and estimated speaker age for English listeners listening to English speakers, followed by Japanese listeners listening to both English and Japanese speakers, with English listeners listening to Japanese speakers coming last. We find that Japanese speakers are estimated to be younger than the English speakers by English listeners, and that both groups of listeners estimate male speakers and speakers with a lower mean fundamental frequency to be older. These results suggest that listeners rely on sociolinguistic information in their speaker age estimations and language familiarity plays a role in their success.

Friends with text as data benefits: Assessing and extending the use of automated text analysis in political science and political psychology

Applications of automated text analysis measuring topics, ideology, sentiment or even personality are booming in fields like political science and political psychology. These developments are to be applauded as they bring about novel insights about politics using new sources of (unstructured) data. However, a divide exists between work in both disciplines using text as data. In this paper we argue in favor of more integration across disciplinary boundaries, structuring our case around four key issues in the research process: (i) sampling text; (ii) authorship as meta data; (iii) pre-processing text; (iv) analyzing text. Along the way we demonstrate that an assessment of speaker characteristics may crucially depend on the text sources under study, and that the use of sentiment words correlates with estimates of policy positions, with implications for interpretation of the latter. As such, this paper contributes to a critical discussion about the merits of automated text analysis methods in political psychology and political science, with an eye towards advancing the considerable potential of text as data in the study of politics.

Global Pitch Characteristics of Tonal and Non-Tonal Language Speakers : Evidence from Mandarin-Korean Bilinguals

본 연구는 제1언어와 제2언어로 산출된 표준중국어와 한국어에 나타나는 발화 음역과 음높이의 음향 특징을 정량적으로 분석하여, (1) 제1언어 발화에 나타나는 성조언어와 비성조언어의 음역과 음높이 특징 및 차이, (2) 비성조언어가 제1언어인 화자가 성조언어를 제2언어로 발화할 때 및 성조언어가 제1언어인 화자가 비성조언어를 제2언어로 발화할 때 나타나는 음역과 음높이 특징을 고찰한다. 22명의 표준중국어-한국어 이중언어화자가 발화한 ‘북풍과 햇님’을 분석한 결과에 따르면, 제1언어로 산출된 표준중국어와 한국어를 비교할 때 전자가 후자에 비하여 발화 음역이 넓거나 높은 음높이가 실현되는 양상이 보이지 않았다. 이는 성조언어가 비성조언어에 비하여 반드시 음역이 전반적으로 상승하거나 확대되는 것은 아님을 시사하며, 성조언어와 비성조언어의 억양 차이를 발화체 전반에 걸쳐 실현되는 음역과 음높이의 차이로 일반화하는 기존 가설에 문제점을 제기한다. 또한 제2언어로 산출된 표준중국어와 한국어를 비교한 결과, 유의미한 음역의 축소는 표준중국어 여성 모어화자가 발화한 한국어에서만 발견되었다. 이에 반해 표준중국어 남성 모어화자와 한국어 남성, 여성 모어화자가 발화한 제2언어에서는 제1언어와의 큰 차이가 나타나지 않았다. 더 나아가 제2언어에서 음역이 조절되는 방식은 화자의 성별이 중요한 영향을 미친다는 점을 발견하였다. 이러한 결과는 제2언어의 숙달도 및 제2언어화자의 메타언어적 지식이 발화의 음역 산출에 영향을 준다는 점을 시사한다.

Morphological characteristics of male and female hypopharynx: A magnetic resonance imaging-based study.

Studies with three-dimensional (3D) vocal tract visualization using magnetic resonance imaging (MRI) have suggested that hypopharyngeal cavities, i.e., laryngeal cavity and bilateral piriform fossa, may be the acoustic loci to express speaker characteristics in male speech sounds. Previous studies mainly investigated the hypopharynx in males but few for females. This study explored the hypopharynx morphological characteristics for the subjects in both genders by MRI. 3D numerical vocal tracts at vowels were reconstructed from the MRI datasets of three male and four female Chinese subjects. Geometrical measurements were conducted for the hypopharyngeal cavities. Morphological observations and statistical analyses revealed both commonalities and differences between the male and female subjects. The laryngeal cavity shapes in females were found similar to males resembling a Helmholtz resonator rather than a simple straight closed tube, and the bilateral piriform fossa cavities showed an asymmetry: the right is longer and wider than the left in both genders. As for the cavity size across vowels, for both the male and female subjects the laryngeal cavity and piriform fossa in /i/ were observed larger than those in /a/. To summarize gender characteristics, the female subjects were characterized by the smaller laryngeal cavity and piriform fossa compared with males.

The Intonation and Pragmatics of Greek wh-Questions.

We experimentally tested three hypotheses regarding the pragmatics of two tunes (one high-ending, one flat-ending) used with Greek wh-questions: (a) the high-ending tune is associated with information-seeking questions, while the flat-ending tune is also appropriate when wh-questions are not information-seeking, in which case their function can instead be akin to that of a statement; (b) the high-ending tune is more polite, and (c) more appropriate for contexts leading to information-seeking questions. The wh-questions used as experimental stimuli were elicited from four speakers in contexts likely to lead to either information-seeking or non-information-seeking uses. The speakers produced distinct tunes in response to the contexts; acoustic analysis indicates these are best analysed as L*+H L-!H% (rising), and L+H* L-L% (flat). In a perception experiment where participants heard the questions out of context, they chose answers providing information significantly more frequently after high-ending than flat-ending questions, confirming hypothesis (a). In a second experiment testing hypotheses (b) and (c), participants evaluated wh-questions for appropriateness and politeness in information- and non-information-seeking contexts. High-ending questions were rated more appropriate in information-seeking contexts, and more polite independently of context relative to their flat-ending counterparts. Finally, two follow-up experiments showed that the interpretation of the two tunes was not affected by voice characteristics of individual speakers, and confirmed a participant preference for the high-ending tune. Overall, the results support our hypotheses and lead to a compositional analysis of the meaning of the two tunes, while also showing that intonational meaning is determined by both tune and pragmatic context.

Voice Conversion With CycleRNN-Based Spectral Mapping and Finely Tuned WaveNet Vocoder

In this paper, we present a novel framework for a voice conversion (VC) system based on a cyclic recurrent neural network (CycleRNN) and a finely tuned WaveNet vocoder. Even though WaveNet is capable of producing natural speech waveforms when fed with natural speech features, it still suffers from speech quality degradation when fed with oversmoothed features, such as spectral parameters estimated from a statistical model. One way to address this problem is to introduce oversmoothed features while developing a WaveNet model. However, in a VC framework, providing oversmoothed spectral features of a target speaker for WaveNet modeling is not straightforward owing to the difference in the time-sequence alignment from that of a source speaker. To overcome this problem, we propose the use of a cyclic spectral conversion network, i.e., CycleRNN, capable of performing a conversion flow, i.e., source-to-target, and a cyclic flow, i.e., to generate self-predicted target spectra. The CycleRNN spectral model is trained using both conversion and weighted cyclic losses. To finely tune WaveNet, a pretrained multispeaker WaveNet model is optimized using the self-predicted features of the corresponding target speaker of a speaker conversion pair. The experimental results demonstrate that 1) the proposed CycleRNN-based spectral model for WaveNet fine-tuning significantly improves the naturalness of the converted speech waveforms, giving an overall mean opinion score of 3.50; and 2) the proposed model yields the highest speaker conversion accuracy, giving an overall speaker similarity score of 78.33%, which is a significant improvement compared with conventional WaveNet fine-tuning using natural target features.

Análise acústica do ditongo [eI] em inglês e em português por falantes bilíngues

It is said that diphthongs carry potentially useful individualizing information about the speaker and this has been investigated and used in forensic analysis. Also, research projects concerned with bilingualism and differences between L1 and L2 are becoming more and more common contexts for forensic phonetics. These two facts motivated this study into bilingualism and acoustic analysis, specifically into diphthong production. The diphthong [aI] has been the focus of a great deal of research, but the focus of our research was the diphthong [eI], which presents more complexity due to the acoustic proximity of vowel and glide. This article aims at demonstrating the results of acoustic analyses of [eI] in pairs of Portuguese and English words such as lei/lay and leis/lays, as produced by bilingual speakers. The results show that the comparison of acoustic measures, such as duration and formant frequency, can reveal special characteristics of groups of speakers and idiosyncratic features of individual speakers, in both L1 and L2.

Deep CNNs With Self-Attention for Speaker Identification

Most current works on speaker identification are based on i-vector methods; however, there is a marked shift from the traditional i-vector to deep learning methods, especially in the form of convolutional neural networks (CNNs). Rather than designing features and a subsequent individual classification model, we address the problem by learning features and recognition systems using deep neural networks. Based on the deep convolutional neural network (CNN), this paper presents a novel text-independent speaker identification method for speaker separation. Specifically, this paper is based on the two representative CNNs, called the visual geometry group (VGG) nets and residual neural networks (ResNets). Unlike prior deep neural network-based speaker identification methods that usually rely on a temporal maximum or average pooling across all time steps to map variable-length utterances to a fixed-dimension vector, this paper equips these two CNNs with a structured self-attention mechanism to learn a weighted average across all time steps. Using the structured self-attention layer with multiple attention hops, the proposed deep CNN network is not only capable of handling variable-length segments but also able to learn speaker characteristics from different aspects of the input sequence. The experimental results on the speaker identification benchmark database, VoxCeleb demonstrate the superiority of the proposed method over the traditional i-vector-based methods and the other strong CNN baselines. In addition, the results suggest that it is possible to cluster unknown speakers using the activation of an upper layer of a pre-trained identification CNN as a speaker embedding vector.

Estimating cognitive load from speech gathered in a complex real-life training exercise

Abstract Speech-enabled applications are becoming prevalent, providing opportunities for real-time detection of speaker characteristics. Estimation of cognitive load from speech is one type of speaker characteristic that can provide insight into the human state in complex, highly dynamic human-machine teaming scenarios and be used to adapt interaction with the user to their current cognitive state. Cognitive load estimation from speech experiments are typically performed on speech gathered in laboratory settings. By contrast, this research is performed on a real-life dataset that was not created for the purpose of cognitive load assessment. Speech was extracted from recordings of a military simulation exercise in which air battle managers communicated with pilots flying simulated aircraft. This paper assesses whether cognitive load can be estimated from speech self-labelled by exercise participants and collected in a realistic setting, and examines how well cognitive load estimation methods translate from the laboratory setting to the real-world. Analysis suggests that participants’ self-assessment of workload at periodic intervals can be used to label speech to create 2-class cognitive load classifiers. The analysis also shows that including some target speaker speech in speaker independent training data results in higher classification accuracy than when classifiers are built solely from speaker dependent data.

Time-varying sinusoidal demodulation for non-stationary modeling of speech

Speech signals contain a fairly rich time-evolving spectral content. Accurate analysis of this time-evolving spectrum is an open challenge in signal processing. Towards this, we visit time-varying sinusoidal modeling of speech and propose an alternate model estimation approach. The estimation operates on the whole signal without any short-time analysis. The approach proceeds by extracting the fundamental frequency sinusoid (FFS) from speech signal. The instantaneous amplitude (IA) of the FFS is used for voiced/unvoiced stream segregation. The voiced stream is then demodulated using a variant of in-phase and quadrature-phase demodulation carried at harmonics of the FFS. The result is a non-parametric time-varying sinusoidal representation, specifically, an additive mixture of quasi-harmonic sinusoids for voiced stream and a wideband mono-component sinusoid for unvoiced stream. The representation is evaluated for analysis-synthesis, and the bandwidth of IA and IF signals are found to be crucial in preserving the quality. Also, the obtained IA and IF signals are found to be carriers of perceived speech attributes, such as speaker characteristics and intelligibility. On comparing the proposed modeling framework with the existing approaches, which operate on short-time segments, improvement is found in simplicity of implementation, objective-scores, and computation time. The listening test scores suggest that the quality preserves naturalness but does not yet beat the state-of-the-art short-time analysis methods. In summary, the proposed representation lends itself for high resolution temporal analysis of non-stationary speech signals, and also allows quality preserving modification and synthesis.

QUANTITATIVE ANALYSIS OF PITCH MOVEMENT IN SPONTANEOUS SPEECH OF REGIONAL RUSSIAN NATIVE SPEAKERS

The paper deals with the intonation of native speakers of several regional variants of Russian. I analyze how simple characteristics of pitch movement are connected to biological sex, age, place of living, and type of spontaneous text. According to my primary hypothesis, the distribution of pitch in “low”, “middle” and “high” parts of the pitch range and the number of movements from one range into other ones can predict the aforementioned characteristics of the recordings as well as reveal the individual characteristics of speakers (Vol’skaya 2014, Fery 2017). I demonstrate that in my model the only significant parameter is the biological sex of the speakers, which means that this distribution is unlikely to be determined by the regional differences. The study has been conducted as a part of the project on statistical analysis of regional variants of Standard Russian and minority languages of Russia.

Style-Code Method for Multi-Style Parametric Text-to-Speech Synthesis

Modern text-to-speech systems generally achieve good intelligibility. The one of the main drawbacks of these systems is the lack of expressiveness in comparison to natural human speech. It is very unpleasant when automated system conveys positive and negative message in completely the same way. The introduction of parametric methods in speech synthesis gave possibility to easily change speaker characteristics and speaking styles. In this paper a simple method for incorporating styles into synthesized speech by using style codes is presented. The proposed method requires just a couple of minutes of target style and moderate amount of neutral speech. It is successfully applied to both hidden Markov models and deep neural networks-based synthesis, giving style code as additional input to the model. Listening tests confirmed that better style expressiveness is achieved by deep neural networks synthesis compared to hidden Markov model synthesis. It is also proved that quality of speech synthesized by deep neural networks in a certain style is comparable with the speech synthesized in neutral style, although the neutral-speech-database is about 10 times bigger. DNN based TTS with style codes are further investigated by comparing the quality of speech produced by single-style modeling and multi-style modeling systems. Objective and subjective measures confirmed that there is no significant difference between these two approaches.