Multiple Acoustic Features Research Articles

Video quality evaluation toward complicated sport activities for clustering analysis

Automatically clustering various sophisticated human activities (e.g., dancing, martial arts, and gymnastics) based on their quality scores is an indispensable technique in physical training, human–computer interaction, etc. Conventionally, many action recognition models are built upon the visual/semantic appearance of human body movements. Recently, due to the introduction of Microsoft Kinect, many skeleton-based human action understanding frameworks have been proposed. In this work, we propose a novel method to cluster the quality of complicated human actions towards contactless operative video reading system (COVRS). More specifically, we first extract the skeleton by leveraging the Kinect, which is subsequently fed into an aggregation deep neural network to extract the deep feature for each human action skeleton. In COVRS, the human hand gesture is an informative clue. Thus, we propose a ranking algorithm to extract the position of human five figures, based on which the deep hand gesture representation is hierarchically learned. Noticeably, it is observable that, the acoustic feature from many human activities also contributes to the quality assessment. We extract multiple acoustic features from the audio associated with each human activity video. Finally, based on the above human skeleton and hand gesture deep features, as well as the shallow acoustic features, we employ a probabilistic model to integrate them for clustering the various human activities using the quality of COVRS. Comprehensive experimental have demonstrated the effectiveness and efficiency of our method. Besides, empirical results have shown that our probabilistic quality model is highly extensible, where additionally visual/acoustic features can be encoded according to different applications.

Comment: Advances in Studying the Vocal Expression of Emotion: Current Contributions and Further Options

I consider the five contributions in this special section as evidence that the research area dealing with the vocal expression of emotion is advancing rapidly, both in terms of the number of pertinent empirical studies and with respect to an ever increasing sophistication of methodology. I provide some suggestions on promising areas for future interdisciplinary research, including work on emotion expression in singing and the potential of vocal symptoms of emotional disorder. As to the popular discussion of the respective role of universality versus language/culture differences, I suggest to move on from exclusively studying the accuracy of recognition in judgment studies to a more differentiated approach adding production aspects, taking into account the multiple vocal and acoustic features that interact to communicate emotion.

Identification of digital voice biomarkers for cognitive health.

Human voice contains rich information. Few longitudinal studies have been conducted to investigate the potential of voice to monitor cognitive health. The objective of this study is to identify voice biomarkers that are predictive of future dementia. Participants were recruited from the Framingham Heart Study. The vocal responses to neuropsychological tests were recorded, which were then diarized to identify participant voice segments. Acoustic features were extracted with the OpenSMILE toolkit (v2.1). The association of each acoustic feature with incident dementia was assessed by Cox proportional hazards models. Our study included 6, 528 voice recordings from 4, 849 participants (mean age 63 ± 15 years old, 54.6% women). The majority of participants (71.2%) had one voice recording, 23.9% had two voice recordings, and the remaining participants (4.9%) had three or more voice recordings. Although all asymptomatic at the time of examination, participants who developed dementia tended to have shorter segments than those who were dementia free (P < 0.001). Additionally, 14 acoustic features were significantly associated with dementia after adjusting for multiple testing (P < 0.05/48 = 1 × 10-3). The most significant acoustic feature was jitterDDP_sma_de (P = 7.9 × 10-7), which represents the differential frame-to-frame Jitter. A voice based linear classifier was also built that was capable of predicting incident dementia with area under curve of 0.812. Multiple acoustic and linguistic features are identified that are associated with incident dementia among asymptomatic participants, which could be used to build better prediction models for passive cognitive health monitoring.

Discriminating between similar alarm calls of contrasting function.

In a pioneering study of signal design, Marler (Marler 1955 Nature176, 6-8. (doi:10.1038/176006a0); Marler 1957 Behaviour11, 13-37. (doi:10.1163/156853956X00066)) argued that the contrasting acoustic design of hawk (seet) and mobbing alarm calls of European passerines reflected their contrasting function. Hawk alarms were high-frequency tones, warning conspecifics to flee but making localization difficult for predators, while mobbing calls were broadband and harsh, allowing easy localization and approach. Contrasting signal features are also consistent with signal detection theory. Discriminating these calls quickly is critical for survival, because hawk alarms require immediate escape. These signals should therefore be selected to be easy to discriminate, reducing the trade-off between immediate fleeing to hawk alarms and unnecessary fleeing to mobbing alarms. Despite these expectations, hawk and mobbing alarm calls of superb fairy-wrens, Malurus cyaneus, are surprisingly similar, raising the question of discriminability without contextual cues. We synthesized these calls on computer, made intermediates and used playbacks to test whether calls can be discriminated acoustically, and if so by what features. We found that birds used multiple acoustic features when discriminating calls, allowing fast discrimination despite overlap in individual parameters. We speculate that the similarity of fairy-wren alarm calls could enhance detectability of both signals, while multiple subtle acoustic differences reduce a trade-off with discriminability. This article is part of the theme issue 'Signal detection theory in recognition systems: from evolving models to experimental tests'.

Effect of articulatory and acoustic features on the intelligibility of speech in noise: An articulatory synthesis study

In noisy conditions, speakers involuntarily change their manner of speaking to enhance the intelligibility of their voices. The increased intelligibility of this so-called Lombard speech is enabled by the change of multiple articulatory and acoustic features. While the major features of Lombard speech are well known from previous studies, little is known about their relative contributions to the intelligibility of speech in noise. This study used an analysis-by-synthesis strategy to explore the contributions of multiple of these features. To this end, an articulatory speech synthesizer was used to synthesize the ten German digit words “Null” to “Neun”, for all 16 combinations of four binary features, i.e., modal vs. pressed phonation, normal vs. increased F1 and F2 formant frequencies, normal vs. increased f0 mean and range, and normal vs. increased duration of vowels. Subjects were asked to try to recognize the synthesized words in the presence of strong pink noise and babble noise. Compared to “plain” speech, the word recognition rate was most improved by pressed phonation, followed by an increased f0 mean and f0 range, and increased formant frequencies. Increased duration of vowels slightly reduced the recognition rate for pink noise but had no effect for babble noise.

Performance analysis of multiple aggregated acoustic features for environment sound classification

Accuracy cognition in the complex and dynamic environment plays a pivotal part in artificial intelligence. Accurate classification of acoustic events is one of the foundations of environment acoustic awareness that has a strong correlation with the selected features. In this paper, the objective is to present a performance analysis of the of different acoustic features aggregation schemes on environment sound classification (ESC) tasks to find the best feature aggregate strategies to overcome the challenging problem of elevating the classification accuracy of environment sounds. With a considerable number of experiments, the feature combination including MFCC, Log-mel Spectrogram, Chroma, Spectral Contrast and Tonnetz achieves the state-of-art classification accuracy on the ESC dataset (85.6%) and 93.4% on the UrbanSound8K dataset.

An Investigation of Critical Frequency Sub-bands of Snoring Sounds for OSA Diagnosis.

Obstructive Sleep Apnea (OSA) is a result of upper airway narrowing during sleep. The upper airway characteristics are likely to manifest in the acoustic characteristics of snoring sounds as snoring is a result of upper airway structure vibrations. In previous studies, researchers have used different regions of the frequency spectrum to diagnose OSA and determine sites of obstruction as well. However, there is no agreement among researchers about the frequency ranges critical for OSA diagnosis. This paper provides the results of a study of snore sound based OSA diagnosis performance using a multiple acoustic features and multiple classifiers. The results of the study may provide useful insights for researchers to identify frequency sub-bands critical for OSA diagnosis.

Robust singer identification of Indian playback singers

Singing voice analysis has been a topic of research to assist several applications in the domain of music information retrieval system. One such major area is singer identification (SID). There has been enormous increase in production of movies and songs in Bollywood industry over the last 50 decades. Surveying this extensive dataset of singers, the paper presents singer identification system for Indian playback singers. Four acoustic features namely—formants, harmonic spectral envelope, vibrato, and timbre—that uniquely describe the singer are extracted from the singing voice segments. Using the combination of these multiple acoustic features, we address the major challenges in SID like the variations in singer’s voice, testing of multilingual songs, and the album effect. Systematic evaluation shows the SID is robust against the variations in singer’s singing style and structure of songs and is effective in identifying the cover songs and singers. The results are investigated on in-house cappella database consisting of 26 singers and 550 songs. By performing dimension reduction of the feature vector and using Support Vector Machine classifier, we achieved an accuracy of 86% using fourfold cross validation process. In addition, performance comparison of the proposed work with other existing approaches reveals the superiority in terms of volume of dataset and song duration.

Homology and Specificity of Natural Sound-Encoding in Human and Monkey Auditory Cortex.

Understanding homologies and differences in auditory cortical processing in human and nonhuman primates is an essential step in elucidating the neurobiology of speech and language. Using fMRI responses to natural sounds, we investigated the representation of multiple acoustic features in auditory cortex of awake macaques and humans. Comparative analyses revealed homologous large-scale topographies not only for frequency but also for temporal and spectral modulations. In both species, posterior regions preferably encoded relatively fast temporal and coarse spectral information, whereas anterior regions encoded slow temporal and fine spectral modulations. Conversely, we observed a striking interspecies difference in cortical sensitivity to temporal modulations: While decoding from macaque auditory cortex was most accurate at fast rates (> 30 Hz), humans had highest sensitivity to ~3 Hz, a relevant rate for speech analysis. These findings suggest that characteristic tuning of human auditory cortex to slow temporal modulations is unique and may have emerged as a critical step in the evolution of speech and language.

Singing activity‐driven Arc expression associated with vocal acoustic plasticity in juvenile songbird

Learned vocalization, including birdsong and human speech, is acquired through self‐motivated vocal practice during the sensitive period of vocal learning. The zebra finch (Taeniopygia guttata) develops a song characterized by vocal variability and crystallizes a defined song pattern as adulthood. However, it remains unknown how vocal variability is regulated with diurnal singing during the sensorimotor learning period. Here, we investigated the expression of activity‐dependent neuroplasticity‐related gene Arc during the early plastic song phase to examine its potential association with vocal plasticity. We first confirmed that multiple acoustic features of syllables in the plastic song were dramatically and simultaneously modulated during the first 3 hr of singing in a day and the altered features were maintained until sleep. In a concurrent manner, Arc was intensely induced during morning singing and a subsequent attenuation during afternoon singing in the robust nucleus of the arcopallium (RA) and the interfacial nucleus of the nidopallium (NIf). The singing‐driven Arc expression was not altered by circadian rhythm, but rather reduced during the day as juveniles produced more songs. Song stabilization accelerated by testosterone administration in juveniles was accompanied with attenuation of Arc induction in RA and NIf. In contrast, although early‐deafened birds produced highly unstable song even at adulthood, singing‐driven Arc expression was not different between intact and early‐deafened adults. These results suggest a potential functional link between Arc expression in RA and NIf and vocal plasticity during the sensorimotor phase of song learning. Nonetheless, Arc expression did not reflect the quality of bird's own song or auditory feedback.

T138. ACOUSTIC PATTERNS IN SCHIZOPHRENIA: A SYSTEMATIC REVIEW AND META-ANALYSIS

BackgroundIndividuals with schizophrenia are characterized as presenting atypical voice patterns: poverty of speech, increased pauses, distinctive pitch (mean and variability). Voice atypicalities may play a role in the social impairment experienced by patients, and could constitute a window into motor, cognitive, emotional and social components of the disorder. Indeed, they have already been generally associated with negative symptoms. However, the state of the evidence for atypical voice patterns and their relation to clinical features is uncertain. Studies using clinical rating scales indicate that voice alterations are severe across many voice properties. In contrast, quantitative acoustic studies seem to have found less robust and more variable results limited to specific features. We therefore systematically reviewed the literature quantifying acoustic patterns in schizophrenia, and performed a meta-analysis of the evidence. We aimed at identifying evidence for acoustic markers of schizophrenia and its clinical features, needs for further research and barriers to collective advancements on these issues.MethodsWe adopted the “PRISMA Statement” guidelines for transparent reporting of a systematic review. The literature search was conducted on Pubmed and Google Scholar (details and pre-registration at https://goo.gl/H1yDpm). Study selection was conducted according to the following inclusion criteria: (a) empirical study, (b) quantification of acoustic features in the vocal production of participants with schizophrenia, (c) sample including at least two individuals with schizophrenia, (d) inclusion of a comparison group, or an assessment of variation in acoustic features in relation to severity of clinical features. We identified 54 studies as eligible for inclusion and contacted all authors to obtain missing estimates and individual-level data, where possible. 34 studies availed enough information to be included in a meta-analysis. The meta-analysis consisted of mixed effects regression models, one per each relevant acoustic feature.ResultsOf the 37 authors contacted, 59% responded and 5% provided at least some of the requested data. Chief reasons of denials were: i) data loss (n = 8), ii) effort required (n = 5), iii) ethical concerns with data sharing (n = 1). On the results available we found significant meta-analytic effects of schizophrenia in percentage of spoken time (n = 6, d = -1.16, 95% CIs: -2.06 -0.27) and proportion of pauses (n = 5, d = 0.56, 95% CIs: 0.15 0.96). After controlling for influential studies, we found significant differences also in pitch mean (n = 5, d = 0.40, 95% CIs: 0.12 0.68) and pitch variability (n = 6, d = -0.46, 95% CIs: -0.70 -0.23). No effects were found for pause duration (n = 7), speech rate (n = 9), speech duration (n = 5) and pitch intensity (n = 5). We found evidence for publication bias for studies investigating pause duration and pitch variability.Key concerns on the meta-analysis are: i) small sample sizes, ii) heterogeneity of task and acoustic processing methods, iii) lack of demographic and clinical individual-level data necessary to control for confounds (e.g. medication and relation to clinical features).DiscussionWe found clear effects of increased pause behavior in schizophrenia and less clear effects of pitch. However, the magnitude of these abnormalities is limited and contrast with the large effect sizes reported by studies using clinical rating scales. Future research should focus on larger sample sizes, systematic assessment of multiple acoustic features and multiple speech tasks, standardized acoustic processing methods, and individual level data available. More reflection is needed on how to make data sharing possible within privacy and ethical constraints.

Chronset: An automated tool for detecting speech onset

The analysis of speech onset times has a longstanding tradition in experimental psychology as a measure of how a stimulus influences a spoken response. Yet the lack of accurate automatic methods to measure such effects forces researchers to rely on time-intensive manual or semiautomatic techniques. Here we present Chronset, a fully automated tool that estimates speech onset on the basis of multiple acoustic features extracted via multitaper spectral analysis. Using statistical optimization techniques, we show that the present approach generalizes across different languages and speaker populations, and that it extracts speech onset latencies that agree closely with those from human observations. Finally, we show how the present approach can be integrated with previous work (Jansen & Watter Behavior Research Methods, 40:744–751, 2008) to further improve the precision of onset detection. Chronset is publicly available online at www.bcbl.eu/databases/chronset.

Striking differences in the loud calls of howler monkey sister species (Alouatta pigra and A. palliata).

Comparing vocalizations across species is useful for understanding acoustic variation at mechanistic and evolutionary levels. Here, we take advantage of the divergent vocalizations of two closely related howler monkey species (Alouatta pigra and A. palliata) to better understand vocal evolution. In addition to comparing multiple acoustic and temporal features of roars and the calling bouts in which they are produced, we tested several predictions. First, A. pigra should have roars with lower fundamental frequency and lower formant dispersion because they are larger than A. palliata and have a larger hyoid apparatus. Second, A. pigra should have faster calling rates, longer roars, longer bouts, and exaggerated call features linked to vocal effort (e.g., nonlinear phenomena and emphasized frequencies) because they are the more aggressive species during intergroup encounters. We found significant interspecific differences supporting our predictions in every tested parameter of roars and bouts, except for roar duration and barking rate. Stepwise discriminant function analyses identified the best features for differentiating roars (acoustic features: formant dispersion followed by highest frequency; temporal features: longest syllable duration followed by number of syllables). Although resembling each other more than they resemble South American howler monkeys, our comparison revealed striking differences in the vocalizations of the two Mesoamerican species. While we cannot completely rule out the influence of body size or the environmental conditions in which the two species evolved, vocal differences were likely influenced by sexual selection. The exaggerated roars and intense calling patterns in A. pigra seem more suitable for intergroup competition, whereas A. palliata calls may be better suited for mate attraction and competition within groups. With interspecific acoustic differences quantified, we will now be able to examine how vocalizations contribute to the evolutionary dynamics of the A. palliata × A. pigra hybrid zone in southern Mexico. Am. J. Primatol. 78:755-766, 2016. © 2016 Wiley Periodicals, Inc.

Distributed acoustic cues for caller identity in macaque vocalization.

Individual primates can be identified by the sound of their voice. Macaques have demonstrated an ability to discern conspecific identity from a harmonically structured ‘coo’ call. Voice recognition presumably requires the integrated perception of multiple acoustic features. However, it is unclear how this is achieved, given considerable variability across utterances. Specifically, the extent to which information about caller identity is distributed across multiple features remains elusive. We examined these issues by recording and analysing a large sample of calls from eight macaques. Single acoustic features, including fundamental frequency, duration and Weiner entropy, were informative but unreliable for the statistical classification of caller identity. A combination of multiple features, however, allowed for highly accurate caller identification. A regularized classifier that learned to identify callers from the modulation power spectrum of calls found that specific regions of spectral–temporal modulation were informative for caller identification. These ranges are related to acoustic features such as the call’s fundamental frequency and FM sweep direction. We further found that the low-frequency spectrotemporal modulation component contained an indexical cue of the caller body size. Thus, cues for caller identity are distributed across identifiable spectrotemporal components corresponding to laryngeal and supralaryngeal components of vocalizations, and the integration of those cues can enable highly reliable caller identification. Our results demonstrate a clear acoustic basis by which individual macaque vocalizations can be recognized.

The effect of audiovisual and binaural listening on the acceptable noise level (ANL): establishing an ANL conceptual model.

Little is known regarding the acoustic features of a stimulus used by listeners to determine the acceptable noise level (ANL). Features suggested by previous research include speech intelligibility (noise is unacceptable when it degrades speech intelligibility to a certain degree; the intelligibility hypothesis) and loudness (noise is unacceptable when the speech-to-noise loudness ratio is poorer than a certain level; the loudness hypothesis). The purpose of the study was to investigate if speech intelligibility or loudness is the criterion feature that determines ANL. To achieve this, test conditions were chosen so that the intelligibility and loudness hypotheses would predict different results. In Experiment 1, the effect of audiovisual (AV) and binaural listening on ANL was investigated; in Experiment 2, the effect of interaural correlation (ρ) on ANL was examined. A single-blinded, repeated-measures design was used. Thirty-two and twenty-five younger adults with normal hearing participated in Experiments 1 and 2, respectively. In Experiment 1, both ANL and speech recognition performance were measured using the AV version of the Connected Speech Test (CST) in three conditions: AV-binaural, auditory only (AO)-binaural, and AO-monaural. Lipreading skill was assessed using the Utley lipreading test. In Experiment 2, ANL and speech recognition performance were measured using the Hearing in Noise Test (HINT) in three binaural conditions, wherein the interaural correlation of noise was varied: ρ = 1 (N(o)S(o) [a listening condition wherein both speech and noise signals are identical across two ears]), -1 (NπS(o) [a listening condition wherein speech signals are identical across two ears whereas the noise signals of two ears are 180 degrees out of phase]), and 0 (N(u)S(o) [a listening condition wherein speech signals are identical across two ears whereas noise signals are uncorrelated across ears]). The results were compared to the predictions made based on the intelligibility and loudness hypotheses. The results of the AV and AO conditions appeared to support the intelligibility hypothesis due to the significant correlation between visual benefit in ANL (AV re: AO ANL) and (1) visual benefit in CST performance (AV re: AO CST) and (2) lipreading skill. The results of the N(o)S(o), NπS(o), and N(u)S(o) conditions negated the intelligibility hypothesis because binaural processing benefit (NπS(o) re: N(o)S(o), and N(u)S(o) re: N(o)S(o)) in ANL was not correlated to that in HINT performance. Instead, the results somewhat supported the loudness hypothesis because the pattern of ANL results across the three conditions (N(o)S(o) ≈ NπS(o) ≈ N(u)S(o) ANL) was more consistent with what was predicted by the loudness hypothesis (N(o)S(o) ≈ NπS(o) < N(u)S(o) ANL) than by the intelligibility hypothesis (NπS(o) < N(u)S(o) < N(o)S(o) ANL). The results of the binaural and monaural conditions supported neither hypothesis because (1) binaural benefit (binaural re: monaural) in ANL was not correlated to that in speech recognition performance, and (2) the pattern of ANL results across conditions (binaural < monaural ANL) was not consistent with the prediction made based on previous binaural loudness summation research (binaural ≥ monaural ANL). The study suggests that listeners may use multiple acoustic features to make ANL judgments. The binaural/monaural results showing that neither hypothesis was supported further indicate that factors other than speech intelligibility and loudness, such as psychological factors, may affect ANL. The weightings of different acoustic features in ANL judgments may vary widely across individuals and listening conditions.

Deep Belief Networks Based Voice Activity Detection

Fusing the advantages of multiple acoustic features is important for the robustness of voice activity detection (VAD). Recently, the machine-learning-based VADs have shown a superiority to traditional VADs on multiple feature fusion tasks. However, existing machine-learning-based VADs only utilize shallow models, which cannot explore the underlying manifold of the features. In this paper, we propose to fuse multiple features via a deep model, called deep belief network (DBN). DBN is a powerful hierarchical generative model for feature extraction. It can describe highly variant functions and discover the manifold of the features. We take the multiple serially-concatenated features as the input layer of DBN, and then extract a new feature by transferring these features through multiple nonlinear hidden layers. Finally, we predict the class of the new feature by a linear classifier. We further analyze that even a single-hidden-layer-based belief network is as powerful as the state-of-the-art models in the machine-learning-based VADs. In our empirical comparison, ten common features are used for performance analysis. Extensive experimental results on the AURORA2 corpus show that the DBN-based VAD not only outperforms eleven referenced VADs, but also can meet the real-time detection demand of VAD. The results also show that the DBN-based VAD can fuse the advantages of multiple features effectively.

Intensity and directionality of bat echolocation signals.

The paper reviews current knowledge of intensity and directionality of bat echolocation signals. Recent studies have revealed that echolocating bats can be much louder than previously believed. Bats previously dubbed “whispering” can emit calls with source levels up to 110 dB SPL at 10 cm and the louder open space hunting bats have been recorded at above 135 dB SPL. This implies that maximum emitted intensities are generally 30 dB or more above initial estimates. Bats' dynamic control of acoustic features also includes the intensity and directionality of their sonar calls. Aerial hawking bats will increase signal directionality in the field along with intensity thus increasing sonar range. During the last phase of prey pursuit, vespertilionid bats broaden their echolocation beam considerably, probably to counter evasive maneuvers of eared prey. We highlight how multiple call parameters (frequency, duration, intensity, and directionality of echolocation signals) in unison define the search volume probed by bats and in turn how bats perceive their surroundings. Small changes to individual parameters can, in combination, drastically change the bat's perception, facilitating successful navigation and food acquisition across a vast range of ecological niches. To better understand the function of echolocation in the natural habitat it is critical to determine multiple acoustic features of the echolocation calls. The combined (interactive) effects, not only of frequency and time parameters, but also of intensity and directionality, define the bat's view of its acoustic scene.

Perceptual Structure of Adductor Spasmodic Dysphonia and Its Acoustic Correlates

To examine the perceptual structure of voice attributes in adductor spasmodic dysphonia (ADSD) before and after botulinum toxin treatment and identify acoustic correlates of underlying perceptual factors. Reliability of perceptual judgments is considered in detail. Pre- and posttreatment trial with comparison to healthy controls, using single-blind randomized listener judgments of voice qualities, as well as retrospective comparison with acoustic measurements. Oral readings were recorded from 42 ADSD speakers before and after treatment as well as from their age- and sex-matched controls. Experienced judges listened to speech samples and rated attributes of overall voice quality, breathiness, roughness, and brokenness, using computer-implemented visual analog scaling. Data were adjusted for regression to the mean and submitted to principal components factor analysis. Acoustic waveforms, extracted from the reading samples, were analyzed and measurements correlated with perceptual factor scores. Four reliable perceptual variables of ADSD voice were effectively reduced to two underlying factors that corresponded to hyperadduction, most strongly associated with roughness, and hypoadduction, most strongly associated with breathiness. After treatment, the hyperadduction factor improved, whereas the hypoadduction factor worsened. Statistically significant (P<0.01) correlations were observed between perceived roughness and four acoustic measures, whereas breathiness correlated with aperiodicity and cepstral peak prominence (CPPs). This study supported a two-factor model of ADSD, suggesting perceptual characterization by both hyperadduction and hypoadduction before and after treatment. Responses of the factors to treatment were consistent with previous research. Correlations among perceptual and acoustic variables suggested that multiple acoustic features contributed to the overall impression of roughness. Although CPPs appears to be a partial correlate of perceived breathiness, a physical basis of this percept remained less clear.

Combining formant frequency based on variable order LPC coding with acoustic features for TIMIT phone recognition

Combination of multiple acoustic features has great potential to improve Automatic Speech Recognition (ASR) accuracy. Our contribution in this research was to investigate one novel method when using voiced formants' features in combination with standard MFCC features in order to enhance TIMIT phone recognition. These voiced features provide accurate formants frequencies using a Variable Order LPC Coding (VO-LPC) algorithm that was combined with continuity constraints. The overall estimating formants were concatenated with MFCC features when a voiced frame could be detected. For feature-level combination, Heteroscedastic Linear Discriminant Analysis (HLDA) based approach had been used successfully to find an optimal linear combination of successive vectors of a single feature stream. A series of experiments on phone recognition speaker-independent continuous-speech had been carried out using a subset of the large read-speech TIMIT phone corpus. Hidden Markov Model Toolkit (HTK) was also used throughout all carried experiments. Using such feature level combination, optimized mixture splitting and a bigram language model, a detailed analysis on this combination performance was discussed for Context-Independent (CI) and Context-Dependent (CD) Hidden Markov Models (HMM). Experimental results from our proposed procedure showed that phone error rate was successfully decreased by about 3%. At phonetic level group, an increase of 8% and of 10% was observed respectively for vowel and liquid group. These results proved clear phone enhancement regarding existing state of the art.

Perception of the proximal voice features for emotional speech

The expression of emotions results in a number of changes to the acoustic signal. Considering that prosodic events may be encoded into the speech signal using multiple acoustic features, many acoustic variations may be redundant for listeners. An understanding of the acoustic changes that listeners can reliably use to differentiate emotional speech will help in the modeling emotions, for example, from a Brunswikian perspective. In the Brunswikian model of emotion communication, the acoustic features are represented as “distal cues” and the subjective perception of these changes are represented as “proximal percepts.” Few attempts have been made to understand the proximal percepts of emotional speech in relation to the acoustic properties they represent. In this study, 19 listeners were asked to rate a set of 160 samples (2 sentences expressed in 8 emotions by 10 speakers) using 12 visual-analog scales (8 prosodic scales and 4 emotion dimensions). Results showed high reliability for all scales except roughness, articulation, and intonation. The emotions were especially differentiated on the dimensions of loudness, sharpness, and speech rate. Regression analysis was used to determine the underlying acoustic cues represented by these scales. The results of this analysis will be discussed.