Representation Of Acoustic Features Research Articles

Query-by-Example Spoken Term Detection for Zero-Resource Languages Using Heuristic Search

Query-by-Example spoken content retrieval is a demanding and challenging task when a large volume of spoken content is piled up in the repositories without annotation. In the absence of annotation, spoken content retrieval is achieved by capturing the similarities between the query and spoken terms from the acoustic feature representation itself. Dynamic Time Warping (DTW) centric techniques identify the optimal alignment between the acoustic feature representations and capture the similarities between query and spoken terms. Despite feasibility, the DTW-centric techniques produce a lot of false alarms due to the variabilities that exist in natural speech and degrade the performance. In the proposed approach, the variability challenges are addressed in two stages. At first, the speaker-independent acoustic feature representation was obtained from the deep convolutional neural networks that reduce the speaker variabilities. In the second stage, the similarities between the query and spoken term were captured using the heuristic search method. The proposed approach was compared with other state-of-the-art methods using Microsoft Low-Resource Language speech corpus. A 3% improvement and 32% reduction in the hit and false alarm ratio were achieved across languages.

Deep Learning-Based Acoustic Feature Representations for Dysarthric Speech Recognition

Deep Learning-Based Acoustic Feature Representations for Dysarthric Speech Recognition

MEG Activity in Visual and Auditory Cortices Represents Acoustic Speech-Related Information during Silent Lip Reading.

Speech is an intrinsically multisensory signal, and seeing the speaker’s lips forms a cornerstone of communication in acoustically impoverished environments. Still, it remains unclear how the brain exploits visual speech for comprehension. Previous work debated whether lip signals are mainly processed along the auditory pathways or whether the visual system directly implements speech-related processes. To probe this, we systematically characterized dynamic representations of multiple acoustic and visual speech-derived features in source localized MEG recordings that were obtained while participants listened to speech or viewed silent speech. Using a mutual-information framework we provide a comprehensive assessment of how well temporal and occipital cortices reflect the physically presented signals and unique aspects of acoustic features that were physically absent but may be critical for comprehension. Our results demonstrate that both cortices feature a functionally specific form of multisensory restoration: during lip reading, they reflect unheard acoustic features, independent of co-existing representations of the visible lip movements. This restoration emphasizes the unheard pitch signature in occipital cortex and the speech envelope in temporal cortex and is predictive of lip-reading performance. These findings suggest that when seeing the speaker’s lips, the brain engages both visual and auditory pathways to support comprehension by exploiting multisensory correspondences between lip movements and spectro-temporal acoustic cues.

An Overview of Spoof Detection in ASV Systems

In current years, voice based application are used broadly in varied applications for speaker recognition. By and by, there is a wide work in the investigation of parodying and against mocking for Automatic Speaker Verification (ASV) framework. The current advancement within the ASV system ends up interest to secure these voice biometric systems for existent world applications. This paper provides the literature of spoofing detection, novel acoustic feature representations, deep learning, end-to-end systems, etc. Moreover, it conjointly summaries previous studies of spoofing attacks with stress on SS, VC, and replay alongside recent efforts to develop countermeasures for spoof speech detection and speech sound disorder tasks.

Temporal Encoding is Required for Categorization, But Not Discrimination.

Core auditory cortex (AC) neurons encode slow fluctuations of acoustic stimuli with temporally patterned activity. However, whether temporal encoding is necessary to explain auditory perceptual skills remains uncertain. Here, we recorded from gerbil AC neurons while they discriminated between a 4-Hz amplitude modulation (AM) broadband noise and AM rates >4Hz. We found a proportion of neurons possessed neural thresholds based on spike pattern or spike count that were better than the recorded session's behavioral threshold, suggesting that spike count could provide sufficient information for this perceptual task. A population decoder that relied on temporal information outperformed a decoder that relied on spike count alone, but the spike count decoder still remained sufficient to explain average behavioral performance. This leaves open the possibility that more demanding perceptual judgments require temporal information. Thus, we asked whether accurate classification of different AM rates between 4 and 12Hz required the information contained in AC temporal discharge patterns. Indeed, accurate classification of these AM stimuli depended on the inclusion of temporal information rather than spike count alone. Overall, our results compare two different representations of time-varying acoustic features that can be accessed by downstream circuits required for perceptual judgments.

The Effectiveness of Unsupervised Subword Modeling With Autoregressive and Cross-Lingual Phone-Aware Networks

This study addresses unsupervised subword modeling, i.e., learning acoustic feature representations that can distinguish between subword units of a language. We propose a two-stage learning framework that combines self-supervised learning and cross-lingual knowledge transfer. The framework consists of autoregressive predictive coding (APC) as the front-end and a cross-lingual deep neural network (DNN) as the back-end. Experiments on the ABX subword discriminability task conducted with the Libri-light and ZeroSpeech 2017 databases showed that our approach is competitive or superior to state-of-the-art studies. Comprehensive and systematic analyses at the phoneme- and articulatory feature (AF)-level showed that our approach was better at capturing diphthong than monophthong vowel information, while also differences in the amount of information captured for different types of consonants were observed. Moreover, a positive correlation was found between the effectiveness of the back-end in capturing a phoneme's information and the quality of the cross-lingual phone labels assigned to the phoneme. The AF-level analysis together with t-SNE visualization results showed that the proposed approach is better than MFCC and APC features in capturing manner and place of articulation information, vowel height, and backness information. Taken together, the analyses showed that the two stages in our approach are both effective in capturing phoneme and AF information. Nevertheless, monophthong vowel information is less well captured than consonant information, which suggests that future research should focus on improving capturing monophthong vowel information.

Structured Gaussian Process Regression of Music Mood

Modeling the music mood has wide applications in music categorization, retrieval, and recommendation systems; however, it is challenging to computationally model the affective content of music due to its subjective nature. In this work, a structured regression framework is proposed to model the valence and arousal mood dimensions of music using a single regression model at a linear computational cost. To tackle the subjectivity phenomena, a confidence-interval based estimated consensus is computed by modeling the behavior of various annotators (e.g. biased, adversarial) and is shown to perform better than using the average annotation values. For a compact feature representation of music clips, variational Bayesian inference is used to learn the Gaussian mixture model representation of acoustic features and chord-related features are used to improve the valence estimation by probing the chord progressions between chroma frames. The dimensionality of features is further reduced using an adaptive version of kernel PCA. Using an efficient implementation of twin Gaussian process for structured regression, the proposed work achieves a significant improvement in R2 for arousal and valence dimensions relative to state-of-the-art techniques on two benchmark datasets for music mood estimation.

Recognition of emotion from speech using evolutionary cepstral coefficients

An optimal representation of acoustic features is an ongoing challenge in automatic speech emotion recognition research. In this study, we proposed Cepstral coefficients based on evolutionary filterbanks as emotional features. It is difficult to guarantee that an individual optimized filterbank provides the best representation for emotion classification. Consequently, we employed six HMM-based binary classifiers that used a specific filterbank, which was optimized by a genetic algorithm to categorize the data into seven emotion classes. These optimized classifiers were applied in a hierarchical manner and outperformed conventional Mel Frequency Cepstral Coefficients in terms of overall emotion classification accuracy. The proposed method using evolutionary-based Cepstral coefficients achieved a weighted average recall of 87.29% on the Berlin database while the same approach but using conventional Cepstral features achieved only 79.63%.

Learning Continuous Facial Actions From Speech for Real-Time Animation

Speech conveys not only the verbal communication, but also emotions, manifested as facial expressions of the speaker. In this article, we present deep learning frameworks that directly infer facial expressions from just speech signals. Specifically, the time-varying contextual non-linear mapping between audio stream and micro facial movements is realized by our proposed recurrent neural networks to drive a 3D blendshape face model in real-time. Our models not only activate appropriate facial action units (AUs), defined as 3D expression blendshapes in the FaceWarehouse database, to depict different utterance generating actions in the form of lip movements, but also, without any assumption, automatically estimate emotional intensity of the speaker and reproduces her ever-changing affective states by adjusting strength of related facial unit activations. In the baseline models, conventional handcrafted acoustic features are utilized to predict facial actions. Furthermore, we show that it is more advantageous to learn meaningful acoustic feature representation from speech spectrograms with convolutional nets, which subsequently improves the accuracy of facial action synthesis. Experiments on diverse audiovisual corpora of different actors across a wide range of facial actions and emotional states show promising results of our approaches. Being speaker-independent, our generalized models are readily applicable to various tasks in human-machine interaction and animation.

Affective Latent Representation of Acoustic and Lexical Features for Emotion Recognition.

In this paper, we propose a novel emotion recognition method based on the underlying emotional characteristics extracted from a conditional adversarial auto-encoder (CAAE), in which both acoustic and lexical features are used as inputs. The acoustic features are generated by calculating statistical functionals of low-level descriptors and by a deep neural network (DNN). These acoustic features are concatenated with three types of lexical features extracted from the text, which are a sparse representation, a distributed representation, and an affective lexicon-based dimensions. Two-dimensional latent representations similar to vectors in the valence-arousal space are obtained by a CAAE, which can be directly mapped into the emotional classes without the need for a sophisticated classifier. In contrast to the previous attempt to a CAAE using only acoustic features, the proposed approach could enhance the performance of the emotion recognition because combined acoustic and lexical features provide enough discriminant power. Experimental results on the Interactive Emotional Dyadic Motion Capture (IEMOCAP) corpus showed that our method outperformed the previously reported best results on the same corpus, achieving 76.72% in the unweighted average recall.

Advances in anti-spoofing: from the perspective of ASVspoof challenges

Abstract In recent years, automatic speaker verification (ASV) is used extensively for voice biometrics. This leads to an increased interest to secure these voice biometric systems for real-world applications. The ASV systems are vulnerable to various kinds of spoofing attacks, namely, synthetic speech (SS), voice conversion (VC), replay, twins, and impersonation. This paper provides the literature review of ASV spoof detection, novel acoustic feature representations, deep learning, end-to-end systems, etc. Furthermore, the paper also summaries previous studies of spoofing attacks with emphasis on SS, VC, and replay along with recent efforts to develop countermeasures for spoof speech detection (SSD) task. The limitations and challenges of SSD task are also presented. While several countermeasures were reported in the literature, they are mostly validated on a particular database, furthermore, their performance is far from perfect. The security of voice biometrics systems against spoofing attacks remains a challenging topic. This paper is based on a tutorial presented at APSIPA Annual Summit and Conference 2017 to serve as a quick start for those interested in the topic.

Audio Data Mining for Anthropogenic Disaster Identification: An Automatic Taxonomy Approach

Disasters are undesirable and often sudden events causing human, material and economic losses, which exceed the coping capability of the affected community or society. In recent years, with significant advancement in information technology, various intelligent systems have been developed to support various aspects of disaster management, including emergency prediction, timely response and aftermath recovery. This paper addresses the anthropogenic disaster identification issue by exploiting ambient sound data. Specifically, a novel and efficient acoustic event classification scheme is proposed, which is based on unsupervised acoustic feature learning and data-driven taxonomy. The proposed framework could accurately identify anthropogenic disaster events, e.g., gun shot, explosion, scream cry, etc. from dynamic audio data. and it consists of three major stages as follows. First, predominant acoustic patterns are characterized by dictionary learning algorithms, which can generate robust acoustic feature representations for recognition under noisy conditions. Second, hazard sound event taxonomy is created by exploiting probabilistic distances between extracted class-wise dictionaries. Finally, taxonomy structure is embedded into hierarchical classification algorithm to improve event identification performance. The Proposed approach is evaluated using real-world dataset with 10 emergency sound categories and 3,275 clips. According to extensive experimental comparisons, proposed approach achieved state-of-the-art performance in anthropogenic disaster identification.

Cortical encoding of speech enhances task-relevant acoustic information.

Speech is the most important signal in our auditory environment, and the processing of speech is highly dependent on context. However, it is unknown how contextual demands influence the neural encoding of speech. Here, we examine the context dependence of auditory cortical mechanisms for speech encoding at the level of the representation of fundamental acoustic features (spectrotemporal modulations) using model-based functional magnetic resonance imaging. We found that the performance of different tasks on identical speech sounds leads to neural enhancement of the acoustic features in the stimuli that are critically relevant to task performance. These task effects were observed at the earliest stages of auditory cortical processing, in line with interactive accounts of speech processing. Our work provides important insights into the mechanisms that underlie the processing of contextually relevant acoustic information within our rich and dynamic auditory environment.

Speech Emotion Recognition with Heterogeneous Feature Unification of Deep Neural Network.

Automatic speech emotion recognition is a challenging task due to the gap between acoustic features and human emotions, which rely strongly on the discriminative acoustic features extracted for a given recognition task. We propose a novel deep neural architecture to extract the informative feature representations from the heterogeneous acoustic feature groups which may contain redundant and unrelated information leading to low emotion recognition performance in this work. After obtaining the informative features, a fusion network is trained to jointly learn the discriminative acoustic feature representation and a Support Vector Machine (SVM) is used as the final classifier for recognition task. Experimental results on the IEMOCAP dataset demonstrate that the proposed architecture improved the recognition performance, achieving accuracy of 64% compared to existing state-of-the-art approaches.

Adaptation of the human auditory cortex to changing background noise

Speech communication in real-world environments requires adaptation to changing acoustic conditions. How the human auditory cortex adapts as a new noise source appears in or disappears from the acoustic scene remain unclear. Here, we directly measured neural activity in the auditory cortex of six human subjects as they listened to speech with abruptly changing background noises. We report rapid and selective suppression of acoustic features of noise in the neural responses. This suppression results in enhanced representation and perception of speech acoustic features. The degree of adaptation to different background noises varies across neural sites and is predictable from the tuning properties and speech specificity of the sites. Moreover, adaptation to background noise is unaffected by the attentional focus of the listener. The convergence of these neural and perceptual effects reveals the intrinsic dynamic mechanisms that enable a listener to filter out irrelevant sound sources in a changing acoustic scene.

Acoustic spectral imaging and transfer learning for reliable bearing fault diagnosis under variable speed conditions

Incipient fault diagnosis of a bearing requires robust feature representation for an accurate condition-based monitoring system. Existing fault diagnosis schemes are mostly confined to manual features and traditional machine learning approaches such as artificial neural networks (ANN) and support vector machines (SVM). These handcrafted features require substantial human expertise and domain knowledge. In addition, these feature characteristics vary with the bearing’s rotational speed. Thus, such methods do not yield the best results under variable speed conditions. To address this issue, this paper presents a reliable fault diagnosis scheme based on acoustic spectral imaging (ASI) of acoustic emission (AE) signals as a precise health state. These health states are further utilized with transfer learning, which is a machine learning technique, which shares knowledge with convolutional neural networks (CNN) for accurate diagnosis under variable operating conditions. In ASI, the amplitudes of the spectral components of the windowed time-domain acoustic emission signal are transformed into spectrum imaging. ASI provides a visual representation of acoustic emission spectral features in images. This ensures enhanced spectral images for transfer learning (TL) testing and training, and thus provides a robust classifier technique with high diagnostic accuracy.

Exploring Deep Spectrum Representations via Attention-Based Recurrent and Convolutional Neural Networks for Speech Emotion Recognition

The automatic detection of an emotional state from human speech, which plays a crucial role in the area of human–machine interaction, has consistently been shown to be a difficult task for machine learning algorithms. Previous work on emotion recognition has mostly focused on the extraction of carefully hand-crafted and highly engineered features. Results from these works have demonstrated the importance of discriminative spatio-temporal features to model the continual evolutions of different emotions. Recently, spectrogram representations of emotional speech have achieved competitive performance for automatic speech emotion recognition (SER). How machine learning algorithms learn the effective compositional spatio-temporal dynamics for SER has been a fundamental problem of deep representations, herein denoted as deep spectrum representations. In this paper, we develop a model to alleviate this limitation by leveraging a parallel combination of attention-based bidirectional long short-term memory recurrent neural networks with attention-based fully convolutional networks (FCN). The extensive experiments were undertaken on the interactive emotional dyadic motion capture (IEMOCAP) and FAU aibo emotion corpus (FAU-AEC) to highlight the effectiveness of our approach. The experimental results indicate that deep spectrum representations extracted from the proposed model are well-suited to the task of SER, achieving a WA of 68.1% and a UA of 67.0% on IEMOCAP, and 45.4% for UA on FAU-AEC dataset. Key results indicate that the extracted deep representations combined with a linear support vector classifier are comparable in performance with eGeMAPS and COMPARE, two standard acoustic feature representations.

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.

Numerical value biases sound localization

Speech recognition starts with representations of basic acoustic perceptual features and ends by categorizing the sound based on long-term memory for word meaning. However, little is known about whether the reverse pattern of lexical influences on basic perception can occur. We tested for a lexical influence on auditory spatial perception by having subjects make spatial judgments of number stimuli. Four experiments used pointing or left/right 2-alternative forced choice tasks to examine perceptual judgments of sound location as a function of digit magnitude (1–9). The main finding was that for stimuli presented near the median plane there was a linear left-to-right bias for localizing smaller-to-larger numbers. At lateral locations there was a central-eccentric location bias in the pointing task, and either a bias restricted to the smaller numbers (left side) or no significant number bias (right side). Prior number location also biased subsequent number judgments towards the opposite side. Findings support a lexical influence on auditory spatial perception, with a linear mapping near midline and more complex relations at lateral locations. Results may reflect coding of dedicated spatial channels, with two representing lateral positions in each hemispace, and the midline area represented by either their overlap or a separate third channel.

Cortical processing of pitch: Model-based encoding and decoding of auditory fMRI responses to real-life sounds

Pitch is a perceptual attribute related to the fundamental frequency (or periodicity) of a sound. So far, the cortical processing of pitch has been investigated mostly using synthetic sounds. However, the complex harmonic structure of natural sounds may require different mechanisms for the extraction and analysis of pitch. This study investigated the neural representation of pitch in human auditory cortex using model-based encoding and decoding analyses of high field (7 T) functional magnetic resonance imaging (fMRI) data collected while participants listened to a wide range of real-life sounds. Specifically, we modeled the fMRI responses as a function of the sounds' perceived pitch height and salience (related to the fundamental frequency and the harmonic structure respectively), which we estimated with a computational algorithm of pitch extraction (de Cheveigné and Kawahara, 2002). First, using single-voxel fMRI encoding, we identified a pitch-coding region in the antero-lateral Heschl's gyrus (HG) and adjacent superior temporal gyrus (STG). In these regions, the pitch representation model combining height and salience predicted the fMRI responses comparatively better than other models of acoustic processing and, in the right hemisphere, better than pitch representations based on height/salience alone. Second, we assessed with model-based decoding that multi-voxel response patterns of the identified regions are more informative of perceived pitch than the remainder of the auditory cortex. Further multivariate analyses showed that complementing a multi-resolution spectro-temporal sound representation with pitch produces a small but significant improvement to the decoding of complex sounds from fMRI response patterns.In sum, this work extends model-based fMRI encoding and decoding methods - previously employed to examine the representation and processing of acoustic sound features in the human auditory system - to the representation and processing of a relevant perceptual attribute such as pitch. Taken together, the results of our model-based encoding and decoding analyses indicated that the pitch of complex real life sounds is extracted and processed in lateral HG/STG regions, at locations consistent with those indicated in several previous fMRI studies using synthetic sounds. Within these regions, pitch-related sound representations reflect the modulatory combination of height and the salience of the pitch percept.