Short Utterances Research Articles

Listener Perception of Monopitch, Naturalness, and Intelligibility for Speakers With Parkinson's Disease.

Given the potential significance of speech naturalness to functional and social rehabilitation outcomes, the objective of this study was to examine the effect of listener perceptions of monopitch on speech naturalness and intelligibility in individuals with Parkinson's disease (PD). Two short utterances were extracted from monologue samples of 16 speakers with PD and 5 age-matched adults without PD. Sixteen listeners evaluated these stimuli for monopitch, speech naturalness and intelligibility using the visual sort and rate method. Naïve listeners can reliably judge monopitch, speech naturalness, and intelligibility with minimal familiarization. While monopitch and speech intelligibility were only moderately correlated, monopitch and speech naturalness were highly correlated. A great deal of attention is currently being paid to improvement of vocal loudness and thus speech intelligibility in PD. Our findings suggest that prosodic characteristics such as monopitch should be explored as adjuncts to this treatment of dysarthria in PD. Development of such prosodic treatments may enhance speech naturalness and thus improve quality of life.

강인한 음성인식을 위한 극점 필터링 및 스케일 정규화를 이용한 켑스트럼 특징 정규화 방식

In this paper, the pole filtering concept is applied to the Mel-frequency cepstral coefficient (MFCC) feature vectors in the conventional cepstral mean normalization (CMN) and cepstral mean and variance normalization (CMVN) frameworks. Additionally, performance of the cepstral mean and scale normalization (CMSN), which uses scale normalization instead of variance normalization, is evaluated in speech recognition experiments in noisy environments. Because CMN and CMVN are usually performed on a per-utterance basis, in case of short utterance, they have a problem that reliable estimation of the mean and variance is not guaranteed. However, by applying the pole filtering and scale normalization techniques to the feature normalization process, this problem can be relieved. Experimental results using Aurora 2 database (DB) show that feature normalization method combining the pole-filtering and scale normalization yields the best improvements.

Quality measures based calibration with duration and noise dependency for speaker recognition

This paper studies the effect of short utterances and noise on the performance of automatic speaker recognition. We focus on calibration aspects, and propose a calibration strategy that uses quality measures to model the calibration parameters. We carry out the proposed calibration by using simple Quality Measure Functions (QMFs) of duration and measured signal-to-noise-ratio from speech segments. We test the effectiveness of the approach using two databases, the development set of the I4U collaboration for the NIST Speaker Recognition Evaluation (SRE) 2012, and the evaluation test material of NIST SRE 2012 itself. In comparison with conventional linear calibration, results show that the proposed QMF approach successfully improves the system performance in terms of both discrimination and calibration.

Speaker Recognition of Noisy Short Utterance Based on Speech Frame Quality Discrimination and Three-stage Classification Model

The noisy short utterance is polluted by noise and corpus is less, so the recognition rate significantly decreased. For improving recognition rate, we proposed the dual information quality discrimination algorithm to classify the speech frames: one is differences detection and discrimination algorithm (DDADA), another is the improved SNR discrimination algorithm (ISNRDA). Based on the above two algorithms, the speech frames are classified to three classes: high quality, medium quality and low quality. We proposed GMM-UBM three-stage classification model, and we combine the dual information quality discrimination algorithm with GMM-UBM three-stage classification model. Experiments show that, the dual discrimination quality algorithms can be more precise to classify speech frame, and combining it with GMM-UBM three-stage classification model can make full use of limited corpus of short utterance and can improve the speaker recognition rate of the noisy short utterance.

Infant Preferences for Structural and Prosodic Properties of Infant‐Directed Speech in the Second Year of Life

While a large literature discusses young infants' preference for an infant‐directed speaking style, few studies have explored preferences after the first year. The present work compares infants' preference for two different properties of IDS speech: prosodic changes (primarily pitch and pitch variability) and structural properties (utterance length; lexical repetition). We found that both 12‐ and 16‐month‐old infants continued to prefer listening to speech with the prosodic properties of IDS, but neither age showed any preference for speech with the lexical repetition and short utterances typical of IDS.

Improving the self-adaptive voice activity detector for speaker verification using map adaptation and asymmetric tapers

This paper brings an improvement of voice activity detection, based on vector quantization and speech enhancement preprocessing (VQ-VAD) proposed recently, and applied to speaker verification system under noisy environment. VQ-VAD is based on computing the likelihood ratio on an utterance-by utterance basis from mel-frequency cepstral coefficients that train speech and non-speech models. Whereas the notion of speech and non-speech segments in speech signal is independent of the speaker. For this, a modified VQ-VAD technique is proposed in this paper, by creating two UBM's for speech and non-speech models, trained from a long utterance-independence model. Then, an adaptation of UBM's models to the short utterance of speaker is performed via MAP adaptation, instead of using VQ models. Mel-frequency cepstral coefficient's were also extracted by using the recently proposed asymmetric tapers instead of the traditional Hamming windowing. Using the GMM---UBM as a baseline system for speaker verification, extensive simulation results were done by adding different noise levels to the clean TIMIT database, characterized by its short training and very short testing utterances. The obtained results show the superiority of the proposed GMM-MAP-VAD approach in adverse conditions. Furthermore a drastic reduction in the EER is observed when using asymmetric tapers.

Frame-by-frame language identification in short utterances using deep neural networks

This work addresses the use of deep neural networks (DNNs) in automatic language identification (LID) focused on short test utterances. Motivated by their recent success in acoustic modelling for speech recognition, we adapt DNNs to the problem of identifying the language in a given utterance from the short-term acoustic features. We show how DNNs are particularly suitable to perform LID in real-time applications, due to their capacity to emit a language identification posterior at each new frame of the test utterance. We then analyse different aspects of the system, such as the amount of required training data, the number of hidden layers, the relevance of contextual information and the effect of the test utterance duration. Finally, we propose several methods to combine frame-by-frame posteriors. Experiments are conducted on two different datasets: the public NIST Language Recognition Evaluation 2009 (3 s task) and a much larger corpus (of 5 million utterances) known as Google 5M LID, obtained from different Google Services. Reported results show relative improvements of DNNs versus the i-vector system of 40% in LRE09 3 second task and 76% in Google 5M LID.

Research on Noise Processing and Speech Frame Classification Model of Noisy Short UtteranceSignal Processing

The noise processing is the key of improving recognition rate for the noisy utterance. While for the short utterance, its corpus is less and small amount of speech data is available for testing and training, so making full use of its limit corpus is the key of improving recognition rate of the short utterance. For the noisy short utterance, the noise processing and making full use of the limit corpus are vital. We proposed noise separation algorithm based on constrained Non-negative matrix factorization (CNMF) to make the noise processing. As making full use of the limit corpus, we proposed the improved SNR discrimination algorithm (ISNRDA) and the differences detection and discrimination algorithm (DDADA), we use the two classification algorithm to estimate the quality of the speech frame, and classify the speech frame. Besides, we combine the above classification result with the GMM-UBM three-stage classification model proposed in this paper, so that we can make full use of the limit corpus of the noisy short utterance. Experiments show that the above algorithms can improve speaker recognition performance of noisy short utterance.

Non-Negative Factor Analysis of Gaussian Mixture Model Weight Adaptation for Language and Dialect Recognition

Recent studies show that Gaussian mixture model (GMM) weights carry less, yet complimentary, information to GMM means for language and dialect recognition. However, state-of-the-art language recognition systems usually do not use this information. In this research, a non-negative factor analysis (NFA) approach is developed for GMM weight decomposition and adaptation. This modeling, which is conceptually simple and computationally inexpensive, suggests a new low-dimensional utterance representation method using a factor analysis similar to that of the i-vector framework. The obtained subspace vectors are then applied in conjunction with i-vectors to the language/dialect recognition problem. The suggested approach is evaluated on the NIST 2011 and RATS language recognition evaluation (LRE) corpora and on the QCRI Arabic dialect recognition evaluation (DRE) corpus. The assessment results show that the proposed adaptation method yields more accurate recognition results compared to three conventional weight adaptation approaches, namely maximum likelihood re-estimation, non-negative matrix factorization, and a subspace multinomial model. Experimental results also show that the intermediate-level fusion of i-vectors and NFA subspace vectors improves the performance of the state-of-the-art i-vector framework especially for the case of short utterances.

A Gaussian Selection Method for Speaker Verification with Short Utterances

Speaker recognition systems frequently use GMM-MAP method for modeling speakers. This method represents the speaker using a Gaussian mixture. However, in this mixture not all Gaussian components are truly representative of the speaker. In order to remove the model redundancy, this work proposes a Gaussian selection method to achieve a new GMM model only with the more representative Gaussian components. The results of speaker verification experiments applying the proposal show a similar performance to the baseline; however, the speaker models used have a reduction of 80% compared to the speaker model used as the baseline. Our proposal was also applied to speaker recognition system with short test signals of 15, 5 and 3 seconds obtaining an improvement in EER of 0.43%, 2.64% and 1.60%, respectively, compared to the baseline. The application of this method in real or embedded speaker verification systems could be very useful for reducing computational and memory cost.

A Gaussian Selection Method for Speaker Verification with Short Utterances

Speaker recognition systems frequently use GMM-MAP method for modeling speakers. This method represents the speaker using a Gaussian mixture. However, in this mixture not all Gaussian components are truly representative of the speaker. In order to remove the model redundancy, this work proposes a Gaussian selection method to achieve a new GMM model only with the more representative Gaussian components. The results of speaker verification experiments applying the proposal show a similar performance to the baseline; however, the speaker models used have a reduction of 80% compared to the speaker model used as the baseline. Our proposal was also applied to speaker recognition system with short test signals of 15, 5 and 3 seconds obtaining an improvement in EER of 0.43%, 2.64% and 1.60%, respectively, compared to the baseline. The application of this method in real or embedded speaker verification systems could be very useful for reducing computational and memory cost.

A Gaussian Selection Method for Speaker Verification with Short Utterances

Speaker recognition systems frequently use GMM-MAP method for modeling speakers. This method represents the speaker using a Gaussian mixture. However, in this mixture not all Gaussian components are truly representative of the speaker. In order to remove the model redundancy, this work proposes a Gaussian selection method to achieve a new GMM model only with the more representative Gaussian components. The results of speaker verification experiments applying the proposal show a similar performance to the baseline; however, the speaker models used have a reduction of 80% compared to the speaker model used as the baseline. Our proposal was also applied to speaker recognition system with short test signals of 15, 5 and 3 seconds obtaining an improvement in EER of 0.43%, 2.64% and 1.60%, respectively, compared to the baseline. The application of this method in real or embedded speaker verification systems could be very useful for reducing computational and memory cost.

Brunei Teachers’ Perspectives on Questioning: Investigating the Opportunities to “Talk” in Mathematics Lessons

A teachers’ practice cannot be characterised by a single lesson, hence comparison is best made with lesson sequences that better sample the diversity of a teacher’s practice. In this study, we video recorded lesson sequences in four Year 8 mathematics classrooms, as well as interviewed each of the four teachers in Brunei Darussalam. Because of our methodology and based on the findings from the richness in the data that was collected, there were some features in the video and interview data that emerged. One of the features is the significant short utterances made by the students as well as their respective teachers, and the extent of the teachers’ own and their students’ questioning behaviours in the lessons as perceived by the teachers themselves during the video-stimulated recall interviews. In the four Brunei classrooms that we studied, most of the lessons were so rushed, the teachers did most of the talking and when teachers and students do interact, it almost always involved faster-paced exchanges between them. Thus, restricting students to single words (“yes” or “no”) or short choral responses. Overall, the findings appear to indicate that short utterances implied that there were less (or even no) opportunities for fuller student participation in classroom discussions.

Forced Choices: Role Play and the Problem of Disappearing Syntax

We begin with the observation that some Korean elementary school children in English class, confined to a single situation and even a single language exponent in a role play, appear to produce far more coherent dialogue than the previous week when they were allowed completely free choices in language. We note that at the same time as their dialogue becomes longer, their sentences become shorter (tall and thinner on the printed page), and we suggest that the grammatical complexity is consciously unpacked as discourse complexity. When we turn to a large corpus of longitudinal and cross sectional data, we see, however, that “thinner” dialogue with short utterances is not always more creative, at least not grammatically. We see that the children are deliberately and consciously choosing more elliptical and less creative expressions when they speak. We generalize from this data to show that the problem is no mere artifact of the data sample. We theorize from this generalization to show that the development of the child’s free will in syntax is related to a long chain of such constraints, leading from instincts to habits to intelligent choices oriented toward the immediate environment to acts of free will which are genuinely voluntary and which admit no sovereign but the child’s developing self. Indeed, we shall argue that the exercise of this sovereignty is precisely how the child develops a self in the first place.

How Do You Say ‘Hello’? Personality Impressions from Brief Novel Voices

On hearing a novel voice, listeners readily form personality impressions of that speaker. Accurate or not, these impressions are known to affect subsequent interactions; yet the underlying psychological and acoustical bases remain poorly understood. Furthermore, hitherto studies have focussed on extended speech as opposed to analysing the instantaneous impressions we obtain from first experience. In this paper, through a mass online rating experiment, 320 participants rated 64 sub-second vocal utterances of the word ‘hello’ on one of 10 personality traits. We show that: (1) personality judgements of brief utterances from unfamiliar speakers are consistent across listeners; (2) a two-dimensional ‘social voice space’ with axes mapping Valence (Trust, Likeability) and Dominance, each driven by differing combinations of vocal acoustics, adequately summarises ratings in both male and female voices; and (3) a positive combination of Valence and Dominance results in increased perceived male vocal Attractiveness, whereas perceived female vocal Attractiveness is largely controlled by increasing Valence. Results are discussed in relation to the rapid evaluation of personality and, in turn, the intent of others, as being driven by survival mechanisms via approach or avoidance behaviours. These findings provide empirical bases for predicting personality impressions from acoustical analyses of short utterances and for generating desired personality impressions in artificial voices.

Improving short utterance i-vector speaker verification using utterance variance modelling and compensation techniques

This paper proposes techniques to improve the performance of i-vector based speaker verification systems when only short utterances are available. Short-length utterance i-vectors vary with speaker, session variations, and the phonetic content of the utterance. Well established methods such as linear discriminant analysis (LDA), source-normalized LDA (SN-LDA) and within-class covariance normalization (WCCN) exist for compensating the session variation but we have identified the variability introduced by phonetic content due to utterance variation as an additional source of degradation when short-duration utterances are used. To compensate for utterance variations in short i-vector speaker verification systems using cosine similarity scoring (CSS), we have introduced a short utterance variance normalization (SUVN) technique and a short utterance variance (SUV) modelling approach at the i-vector feature level. A combination of SUVN with LDA and SN-LDA is proposed to compensate the session and utterance variations and is shown to provide improvement in performance over the traditional approach of using LDA and/or SN-LDA followed by WCCN. An alternative approach is also introduced using probabilistic linear discriminant analysis (PLDA) approach to directly model the SUV. The combination of SUVN, LDA and SN-LDA followed by SUV PLDA modelling provides an improvement over the baseline PLDA approach. We also show that for this combination of techniques, the utterance variation information needs to be artificially added to full-length i-vectors for PLDA modelling.

Reading Comprehension in a Large Cohort of French First Graders from Low Socio-Economic Status Families: A 7-Month Longitudinal Study

BackgroundThe literature suggests that a complex relationship exists between the three main skills involved in reading comprehension (decoding, listening comprehension and vocabulary) and that this relationship depends on at least three other factors orthographic transparency, children’s grade level and socioeconomic status (SES). This study investigated the relative contribution of the predictors of reading comprehension in a longitudinal design (from beginning to end of the first grade) in 394 French children from low SES families.Methodology/Principal findingsReading comprehension was measured at the end of the first grade using two tasks one with short utterances and one with a medium length narrative text. Accuracy in listening comprehension and vocabulary, and fluency of decoding skills, were measured at the beginning and end of the first grade. Accuracy in decoding skills was measured only at the beginning. Regression analyses showed that listening comprehension and decoding skills (accuracy and fluency) always significantly predicted reading comprehension. The contribution of decoding was greater when reading comprehension was assessed via the task using short utterances. Between the two assessments, the contribution of vocabulary, and of decoding skills especially, increased, while that of listening comprehension remained unchanged.Conclusion/SignificanceThese results challenge the ‘simple view of reading’. They also have educational implications, since they show that it is possible to assess decoding and reading comprehension very early on in an orthography (i.e., French), which is less deep than the English one even in low SES children. These assessments, associated with those of listening comprehension and vocabulary, may allow early identification of children at risk for reading difficulty, and to set up early remedial training, which is the most effective, for them.

Multi-resolution time frequency feature and complementary combination for short utterance speaker recognition

A human speaker recognition expert often observes the speech spectrogram in multiple different scales for speaker recognition, especially under the short utterance condition. Inspired by this action, this paper proposes a novel multi-resolution time frequency feature (MRTF) extraction method, which is obtained by performing a 2-Dimensional discrete cosine transform (DCT) in multi-scale on the time frequency spectrogram matrix and then selecting and combining to the final multi-scaled transformed elements. Compared to the traditional Mel-Frequency Cepstral Coefficient (MFCC) feature extraction, the proposed method can make better use of multi-resolution temporal-frequency information. Beyond this, we also proposed three complementary combination strategies of MFCC and MRTF: in feature level, in i-vector level and in score level. Comparing their performance. We found the best results are obtained by combination in i-vector level. In the three NIST 2008 Speaker Recognition Evaluation datasets, the proposed method is the most effective for improving the performance under short utterance than under long utterance. And after the combination, we can achieve an EER of 11.32 % and MinDCF of 0.054 in the 10sec-10sec trials on the male dataset, which is an absolute 3 % improvement of EER than the best reported result in this field.

What Does Silence Signify? Investigating the Rhetoric of Silence in Berghuis v. Thompkins

This article uses the Supreme Court's opinions in Berghuis v. Thompkins (2010) to show the risks of silence in communication. Theories of rhetorical silence celebrate silences' communicative potential; yet, Berghuis v. Thompkins (2010) offers rhetorical studies scholars an opportunity to examine the value of silence in a new situation, custodial interrogation. In Berghuis v. Thompkins (2010) the Court ruled that, despite nearly three hours of silence during police interrogation, Thompkins's two short utterances constituted a confession, and waived his right to remain silent. This article analyzes Thompkins's silence, first, using rhetorical silence literature, which fails to explain why the Court did not recognize Thompkins's silence as communicative. Then, this article employs H. P. Grice's theory of conversational implicature, which examines ambiguous utterances, to explain how the Court interpreted Thompkins's silence. Reading the Court's opinions through Grice's theory of implicature demonstrates that theories of rhetorical silence hide the risks silence poses for some rhetors.

A supervised approach for segmentation of bioacoustics audio recordings

Segmentation is one of the most important tasks in preprocessing audio recordings for species recognition. For examples, bird songs or calls often consist of distinct short utterances. Correctly segmenting such utterances is an essential step in the analysis of bird songs or calls. Energy based time-domain segmentation is commonly used and can be fairly effective when dealing with high signal-to-noise ratio recordings of a single individual. We consider the scenario in which omnidirectional microphone are deployed for round-the-clock in-situ species monitoring. In such scenario, recordings may suffer from two problems: (i) low signal-to-noise ratio and (ii) simultaneous vocalizations of multiple individuals. We are interested in segmentation of such recordings. We propose a framework for supervised time-frequency segmentation of audio recordings. Using manually labeled spectrograms, a classifier is trained to separate vocalization from the background noise in the two dimensional time-frequency domain.