Search

Simultaneous Localization And Mapping (SLAM) is crucial for autonomous navigation. ORB-SLAM is a state-of-the-art Visual SLAM system based on cameras used for self-driving cars. In this paper, we propose a high-performance, energy-efficient, and functionally accurate hardware accelerator for ORB-SLAM, focusing on its most time-consuming stage: Oriented FAST and Rotated BRIEF (ORB) feature extraction. The Rotated BRIEF (rBRIEF) descriptor generation is the main bottleneck in ORB computation, as it exhibits highly irregular access patterns to local on-chip memories causing a high-performance penalty due to bank conflicts. We introduce a technique to find an optimal static pattern to perform parallel accesses to banks based on a genetic algorithm. Furthermore, we propose the combination of an rBRIEF pixel duplication cache, selective ports replication, and pipelining to reduce latency without compromising cost. The accelerator achieves a reduction in energy consumption of 14597× and 9609×, with respect to high-end CPU and GPU platforms, respectively. • Feature extraction takes 60% of the localization execution time in current solutions. • Streaming architecture in which the bottleneck is the computation of rBRIEF. • A genetic algorithm can statically reorder the rBRIEF pattern to reduce the latency. • Optimizations include: pixel cache, selective replication of ports and pipelining. • Search of the most energy efficient factors that fulfill the self-drivings requirements. [ABSTRACT FROM AUTHOR]

Predicting the performance of applications executed on GPUs is a great challenge and is essential for efficient job schedulers. There are different approaches to do this, namely analytical modeling and machine learning (ML) techniques. Machine learning requires large training sets and reliable features, nevertheless it can capture the interactions between architecture and software without manual intervention. In this paper, we compared a BSP-based analytical model to predict the time of execution of kernels executed over GPUs. The comparison was made using three different ML techniques. The analytical model is based on the number of computations and memory accesses of the GPU, with additional information on cache usage obtained from profiling. The ML techniques Linear Regression, Support Vector Machine, and Random Forest were evaluated over two scenarios: first, data input or features for ML techniques were the same as the analytical model and, second, using a process of feature extraction, which used correlation analysis and hierarchical clustering. Our experiments were conducted with 20 CUDA kernels, 11 of which belonged to 6 real-world applications of the Rodinia benchmark suite, and the other were classical matrix-vector applications commonly used for benchmarking. We collected data over 9 NVIDIA GPUs in different machines. We show that the analytical model performs better at predicting when applications scale regularly. For the analytical model a single parameter λ is capable of adjusting the predictions, minimizing the complex analysis in the applications. We show also that ML techniques obtained high accuracy when a process of feature extraction is implemented. Sets of 5 and 10 features were tested in two different ways, for unknown GPUs and for unknown Kernels. For ML experiments with a process of feature extractions, we got errors around 1.54% and 2.71%, for unknown GPUs and for unknown Kernels, respectively. [ABSTRACT FROM AUTHOR]

Partially fake audio, a variant of deep fake that involves manipulating audio utterances through the incorporation of fake or externally-sourced bona fide audio clips, constitutes a growing threat as an audio forgery attack impacting both human and artificial intelligence applications. Researchers have recently developed valuable databases to aid in the development of effective countermeasures against such attacks. While existing countermeasures mainly focus on identifying partially fake audio at the level of entire utterances or segments, this paper introduces a paradigm shift by proposing frame-level systems. These systems are designed to detect manipulated utterances and pinpoint the specific regions within partially fake audio where the manipulation occurs. Our approach leverages acoustic features extracted from large-scale self-supervised pre-training models, delivering promising results evaluated on diverse, publicly accessible databases. Additionally, we study the integration of boundary and deepfake detection systems, exploring their potential synergies and shortcomings. Importantly, our techniques have yielded impressive results. We have achieved state-of-the-art performance on the test dataset of the Track 2 of ADD 2022 challenge with an equal error rate of 4.4%. Furthermore, our methods exhibit remarkable performance in locating manipulated regions in Track 2 of the ADD 2023 challenge, resulting in a final ADD score of 0.6713 and securing the top position. • Simultaneously detecting partially spoofed audio and locating the manipulated regions. • Exploring model integration techniques in identifying fake regions. • Investigating self-supervised pre-training models in partially spoofing detection. [ABSTRACT FROM AUTHOR]

Deep learning methods exhibited significant advantages in mapping highly nonlinear relationships with acceptable computational speed, and have been widely used to predict water quality. However, various model selection and construction methods resulted in differences in prediction accuracy and performance. Hence, a unified deep learning framework for water quality prediction was established in the paper, including data processing module, feature enhancement module, and data prediction module. In the established model, the data processing module based on wavelet transform method was applied to decomposing complex nonlinear meteorology, hydrology, and water quality data into multiple frequency domain signals for extracting self characteristics of data cyclic and fluctuations. The feature enhancement module based on Informer Encoder was used to enhance feature encoding of time series data in different frequency domains to discover global time dependent features of variables. Finally, the data prediction module based on the stacked bidirectional long and short term memory network (SBiLSTM) method was employed to strengthen the local correlation of feature sequences and predict the water quality. The established model framework was applied in Lijiang River in Guilin, China. The maximum relative errors between the predicted and observed values for dissolved oxygen (DO), chemical oxygen demand (COD Mn) were 12.4% and 20.7%, suggesting a satisfactory prediction performance of the established model. The validation results showed that the established model was superior to all other models in terms of prediction accuracy with RMSE values 0.329, 0.121, MAE values 0.217, 0.057, SMAPE values 0.022, 0.063 for DO and COD Mn , respectively. Ablation tests confirmed the necessity and rationality of each module for the established model framework. The established method provided a unified deep learning framework for water quality prediction. [Display omitted] • A deep learning model coupling with data feature extraction and enhancement methods was developed • ESWT module was applied to extract the self characteristics of data • Enhancement module was adopted to enhance the global and local feature of data • The established model improves the accuracy of water quality prediction [ABSTRACT FROM AUTHOR]

Most of the state-of-the-art speaker recognition system use natural speech signal (i.e., real speech, spontaneous speech or contextual speech) from the subjects. In this paper, recognition of a person is attempted from his or her hum with the help of machines. This kind of application can be useful to design person-dependent Query-by-Humming (QBH) system and hence, plays an important role in music information retrieval (MIR) system. In addition, it can be also useful for other interesting speech technological applications such as human-computer interaction, speech prosody analysis of disordered speech, and speaker forensics. This paper develops new feature extraction technique to exploit perceptually meaningful (due to mel frequency warping to imitate human perception process for hearing) phase spectrum information along with magnitude spectrum information from the hum signal. In particular, the structure of state-of-the-art feature set, namely, Mel Frequency Cepstral Coefficients (MFCCs) is modified to capture the phase spectrum information. In addition, a new energy measure , namely, Variable length Teager Energy Operator (VTEO) is employed to compute subband energies of different time-domain subband signals (i.e., an output of 24 triangular-shaped filters used in the mel filterbank). We refer this proposed feature set as MFCC-VTMP (i.e., mel frequency cepstral coefficients to capture perceptually meaningful magnitude and phase information via VTEO)The polynomial classifier (which is in-principle similar to other discriminatively-trained classifiers such as support vector machine (SVM) with polynomial kernel) is used as the basis for all the experiments. The effectiveness of proposed feature set is evaluated and consistently found to be better than MFCCs feature set for several evaluation factors, such as, comparison with other phase-based features, the order of polynomial classifier, person (speaker) modeling approach (such as, GMM-UBM and i -vector), the dimension of feature vector, robustness under signal degradation conditions, static vs. dynamic features, feature discrimination measures and intersession variability. [ABSTRACT FROM AUTHOR]

Microplastics refer to plastic particles measuring less than 5 mm, which has led to serious environmental problem and the detection of these tiny particles is crucial for understanding the corresponding distribution and impact on the marine environment. In this paper, an improved faster region-based convolutional neural network (R–CNN) model was developed for the identification and detection of microplastic particles. In the proposed model, the residual network-50 (ResNet-50) is employed as the backbone with the replacement of the traditional one to enhance the feature extraction capability and the feature pyramid networks (FPN) module is introduced together for solving the multi-scale target detection. By using the improved Faster R–CNN model, the network model performance is enhanced where the average confidence of detecting unique microplastic particles in the marine environment reaches as high as 99%. Moreover, the microparticles mixture was bounded precisely via the predicted bounding boxes without missing detection and wrong detection. In this way, the successful identification of polystyrene microplastic particles from the particles suspension with similar shapes but various conditions of backgrounds, brightness, distributions and object sizes, was achieved by employing the proposed improved Faster R–CNN model, enabling the accurate detection of microplastic particles in marine environment. [Display omitted] • The deep learning method is employed for microparticle particles identification. • Microparticles micrographs can be distinguished with improved Faster R–CNN model. • ResNet-50 and FPN are integrated to enhance feature extraction and detection. • The average confidence of detecting microplastic particles reaches as high as 99%. [ABSTRACT FROM AUTHOR]

Recent works on the vulnerability of automatic speaker verification (ASV) systems confirm that malicious spoofing attacks using synthetic speech can provoke significant increase in false acceptance rate. A reliable detection of synthetic speech is key to develop countermeasure for synthetic speech based spoofing attacks. In this paper, we targeted that by focusing on three major types of artifacts related to magnitude, phase and pitch variation, which are introduced during the generation of synthetic speech. We proposed a new approach to detect synthetic speech using score-level fusion of front-end features namely, constant Q cepstral coefficients (CQCCs), all-pole group delay function (APGDF) and fundamental frequency variation (FFV). CQCC and APGDF were individually used earlier for spoofing detection task and yielded the best performance among magnitude and phase spectrum related features, respectively. The novel FFV feature introduced in this paper to extract pitch variation at frame-level, provides complementary information to CQCC and APGDF. Experimental results show that the proposed approach produces the best stand-alone spoofing detection performance using Gaussian mixture model (GMM) based classifier on ASVspoof 2015 evaluation dataset. An overall equal error rate of 0.05% with a relative performance improvement of 76.19% over the next best-reported results is obtained using the proposed method. In addition to outperforming all existing baseline features for both known and unknown attacks, the proposed feature combination yields superior performance for ASV system (GMM with universal background model/i-vector) integrated with countermeasure framework. Further, the proposed method is found to have relatively better generalization ability when either one or both of copy-synthesized data and limited spoofing data are available a priori in the training pool. [ABSTRACT FROM AUTHOR]

Identifying the family of malware can determine their malicious intent and attack patterns, which helps to efficiently analyze large numbers of malware variants. Methods based on traditional machine learning often require a lot of time and resources in feature engineering. Virtually all existing static analysis methods based on malware visualization are derived from grayscale images, while a single low-order feature representation may be detrimental to discovering hidden features in a malware family. Based on these problems, this paper proposes an effective malware classification framework (MalFCS) based on malware visualization and automated feature extraction. MalFCS includes mainly three modules: malware visualization, feature extraction, and classification. First, we visualize malware binaries as entropy graphs based on structural entropy. Second, we present a feature extractor based on deep convolutional neural networks to extract patterns shared by a family from entropy graphs automatically. Finally, we propose an SVM classifier to classify malware based on the extracted features. We evaluate the proposed MalFCS over two widely studied benchmark datasets, i.e., Malimg and Microsoft. Experimental results show that compared with the state-of-the-art methods, MalFCS can obtain excellent classification performance with accuracy of 0.997 and 1, respectively, achieving the state-of-the-art performance. • MalFCS integrates visualization, automated feature extraction and classification. • Visualizing malware as entropy graphs based on structural entropy. • Using convolutional neural networks to extract family patterns from entropy graphs. • Achieving an accuracy of 0.997 and 1 on Malimg and Microsoft dataset, respectively. [ABSTRACT FROM AUTHOR]

Feature extraction is one of the most important preprocessing steps in predicting the interactions between RNAs and proteins by applying machine learning approaches. Despite many efforts in this area, still, no suitable structural feature extraction tool has been designed. Therefore, an online toolbox, named RPINBASE which can be applied to different scopes of biological applications, is introduced in this paper. This toolbox employs efficient nested queries that enhance the speed of the requests and produces desired features in the form of positive and negative samples. To show the capabilities of the proposed toolbox, the developed toolbox was investigated in the aptamer design problem, and the obtained results are discussed. RPINBASE is an online toolbox and is accessible at http://rpinbase.com. • Generating queryable positive samples from PDB complexes • Generating queryable artificial proposed negative samples (based on family and clan) • Providing a wide range of extracted features for primary and secondary structures of RNA and protein macromolecules. • Selecting efficient and high performance dataset powered by preprocessed and stored data to download feature vectors quickly. • Possibility of sample filtering by PDB information and structural features of macromolecules. [ABSTRACT FROM AUTHOR]

The emergence of numerous genome projects has made the experimental classification of the protein localization almost impossible due to the exponential increase in the number of protein samples. However, most of the applications are merely developed for single-plex and completely ignored the presence of one protein at two or more locations in a cell. In this regard, few attempts were carried out to target Multi-label protein localizations; consequently, undesirable accuracies are achieved. This paper presents a novel approach, in which a discrete feature extraction method is fused with physicochemical properties of amino acids by using Chou's general form of Pseudo Amino Acid Composition. The technique is tested on two benchmark datasets namely: Gpos-mploc and Virus-mPLoc. The empirical results demonstrated that the proposed method yields better results via two examined classifiers i.e. ML-KNN and Rank-SVM. It is established that the proposed model has improved values in all performance measures considered for the comparison. • A computational model is developed for Multi label prokaryotic subcellular localization. • Physiochemical and discrete methods are utilized for features • SMOTE is applied as oversampling technique • Multi-label KNN and Rank SVM are used as classifiers • Obtained quite promising results than existing methods. [ABSTRACT FROM AUTHOR]

Keyphrases of a given document represent its main topic and they are used as a simple method to represent the document. Statistical and graph-based models as unsupervised approaches have been mainly studied. The statistical models have some difficulty in extracting keyphrases from a single document because most statistical ones generally require statistical information from a large corpus. On the other hand, graph-based models can extract keyphrases by only using the information from a single document; nevertheless, they have some drawbacks. The scores of the edges can be biased because a single document does not contain sufficient information to score the edges of a graph and this influences the score of the nodes. In this paper, we propose an effective combination method of a statistical model, C-value method, and a graph-based model to overcome the drawbacks of each model. A new scoring method for keyphrase candidates is developed by the graph-based model and the scores calculated by the new method are applied to the modified C-value method to estimate the final importance scores of the keyphrase candidates. Subsequently, the proposed model is evaluated using two datasets, SemEval 2010 and Inspec, and its results outperformed the state-of-the-art model among unsupervised models and the existing graph-based ranking models. [ABSTRACT FROM AUTHOR]

Long-term exposure to thiabendazole (TBZ) pesticide residues in fruits can have harmful effects on human health. Therefore, this paper proposes a highly sensitive and time-efficient method for the exploration of efficient SERS feature extraction algorithms to establish an optimal quantitative model for the rapid detection of TBZ residues in apples. Specifically, shape-homogeneous Au@Ag nanoparticles (NPs) were developed as label-free substrates, and the optimal detection time of obtaining SERS signals generated by the binding of TBZ to the substrate sample was determined to be 6 min. Recursive Feature Elimination (RFE) and Competitive Adaptive Reweighted Sampling (CARS) were employed for feature extraction algorithms to optimize characteristic peaks and compare them to full spectra variables and extract effective variables on linear (Ridge, PLS) and nonlinear (SVM, RF) models, respectively. Results showed that the RFE-RF (Recursive Feature Elimination-Random Forest) model was more powerful in variable selection and achieved the highest predictive result of TBZ (R P 2 = 0.992, RMSEP = 0.012 μg/mL) and the computed LOD (Limitation of Detection) was 0.1 μg/mL based on PCA (Principal Component Analysis). Comparison with ELISA verified that the designed method is feasible and potentially applicable to guarantee the safety of pesticide residues in fruits. [Display omitted] • The SERS detection time was optimized to be most suitable after the 6th minute. • RFE and CARS were used for feature extraction algorithms to optimize the feature peaks. • Nonlinear algorithms have better predictive power than linear algorithms. • The highest predictions of R2 and RMSEP for RFE-RF were 0.992 and 0.012, respectively. [ABSTRACT FROM AUTHOR]

This paper is about authenticating genuine van Gogh paintings from forgeries. The paintings used in the test in this paper are provided by van Gogh Museum and Kröller-Müller Museum. The authentication process depends on two key steps: feature extraction and outlier detection. In this paper, a geometric tight frame and some simple statistics of the tight frame coefficients are used to extract features from the paintings. Then a forward stage-wise rank boosting is used to select a small set of features for more accurate classification so that van Gogh paintings are highly concentrated towards some center point while forgeries are spread out as outliers. Numerical results show that our method can achieve 86.08% classification accuracy under the leave-one-out cross-validation procedure. Our method also identifies five features that are much more predominant than other features. Using just these five features for classification, our method can give 88.61% classification accuracy which is the highest so far reported in literature. Evaluation of the five features is also performed on two hundred datasets generated by bootstrap sampling with replacement. The median and the mean are 88.61% and 87.77% respectively. Our results show that a small set of statistics of the tight frame coefficients along certain orientations can serve as discriminative features for van Gogh paintings. It is more important to look at the tail distributions of such directional coefficients than mean values and standard deviations. It reflects a highly consistent style in van Gogh's brushstroke movements, where many forgeries demonstrate a more diverse spread in these features. [ABSTRACT FROM AUTHOR]

Recombination hotspots in a genome are unevenly distributed. Hotspots are regions in a genome that show higher rates of meiotic recombinations. Computational methods for recombination hotspot prediction often use sophisticated features that are derived from physico-chemical or structure based properties of nucleotides. In this paper, we propose iRSpot-SF that uses sequence based features which are computationally cheap to generate. Four feature groups are used in our method: k-mer composition, gapped k-mer composition, TF-IDF of k-mers and reverse complement k-mer composition. We have used recursive feature elimination to select 17 top features for hotspot prediction. Our analysis shows the superiority of gapped k-mer composition and reverse complement k-mer composition features over others. We have used SVM with RBF kernel as a classification algorithm. We have tested our algorithm on standard benchmark datasets. Compared to other methods iRSpot-SF is able to produce significantly better results in terms of accuracy, Mathew's Correlation Coefficient and sensitivity which are 84.58%, 0.6941 and 84.57%. We have made our method readily available to use as a python based tool and made the datasets and source codes available at: https://github.com/abdlmaruf/iRSpot-SF. An web application is developed based on iRSpot-SF and freely available to use at: http://irspot.pythonanywhere.com/server.html. • Simple sequence based feature extraction • Effective feature selection • Recombination hotspot prediction • Web application and prediction tool [ABSTRACT FROM AUTHOR]

The paper deals with the automatic analysis of real-life telephone conversations between an agent and a customer of a customer care service ( ccs ). The application domain is the public transportation system in Paris and the purpose is to collect statistics about customer problems in order to monitor the service and decide priorities on the intervention for improving user satisfaction. Of primary importance for the analysis is the detection of themes that are the object of customer problems. Themes are defined in the application requirements and are part of the application ontology that is implicit in the ccs documentation. Due to variety of customer population, the structure of conversations with an agent is unpredictable. A conversation may be about one or more themes. Theme mentions can be interleaved with mentions of facts that are irrelevant for the application purpose. Furthermore, in certain conversations theme mentions are localized in specific conversation segments while in other conversations mentions cannot be localized. As a consequence, approaches to feature extraction with and without mention localization are considered. Application domain relevant themes identified by an automatic procedure are expressed by specific sentences whose words are hypothesized by an automatic speech recognition ( asr ) system. The asr system is error prone. The word error rates can be very high for many reasons. Among them it is worth mentioning unpredictable background noise, speaker accent, and various types of speech disfluencies. As the application task requires the composition of proportions of theme mentions, a sequential decision strategy is introduced in this paper for performing a survey of the large amount of conversations made available in a given time period. The strategy has to sample the conversations to form a survey containing enough data analyzed with high accuracy so that proportions can be estimated with sufficient accuracy. Due to the unpredictable type of theme mentions, it is appropriate to consider methods for theme hypothesization based on global as well as local feature extraction. Two systems based on each type of feature extraction will be considered by the strategy. One of the four methods is novel. It is based on a new definition of density of theme mentions and on the localization of high density zones whose boundaries do not need to be precisely detected. The sequential decision strategy starts by grouping theme hypotheses into sets of different expected accuracy and coverage levels. For those sets for which accuracy can be improved with a consequent increase of coverage a new system with new features is introduced. Its execution is triggered only when specific preconditions are met on the hypotheses generated by the basic four systems. Experimental results are provided on a corpus collected in the call center of the Paris transportation system known as ratp . The results show that surveys with high accuracy and coverage can be composed with the proposed strategy and systems. This makes it possible to apply a previously published proportion estimation approach that takes into account hypothesization errors. [ABSTRACT FROM AUTHOR]

This paper proposes a hybrid refinement scheme for more accurate localization of phonetic boundaries by combining three different post-processing techniques, including statistical correction, fusion, and predictive models. A statistical method based on state-level correction is proposed to improve the segmentation results. Effects of search ranges on the statistical correction process are studied and a state selection scheme is used to enhance the correction results. This paper also examines the effects of time resolution, i.e., stepsize, of acoustic models on the accuracy of segmentation. A multi-resolution fusion process is proposed to further refine the statistically corrected results. Finally, predictive models are designed to improve the segmentation accuracy by incorporating various acoustic features and searching around the preliminary boundary with a smaller stepsize. By applying the hybrid refinement scheme on a well-known corpus, significant improvements of segmentation results in terms of segmentation accuracy with different tolerances, mean absolute error (MAE), and root-mean-square error (RMSE) can be observed. Furthermore, a scenario of cross-corpora segmentation is examined in generating the segmentation results for a new corpus with a small set of labeled data. Experimental results show that the proposed refinement procedure can generate segmentation results comparable to those given by well-trained acoustic models obtained from the new corpus. [ABSTRACT FROM AUTHOR]

In this paper we investigate the audio-visual processing of linguistic prosody, more precisely the detection of word prominence, and examine how the additional visual information can be used to increase the robustness when acoustic background noise is present. We evaluate the detection performance for each modality individually and perform experiments using feature and decision fusion. For the latter we also consider the adaptive fusion with fusion weights adjusted to the current acoustic noise level. Our experiments are based on a corpus with 11 English speakers which contains in addition to the speech signal also videos of the speakers’ heads. From the acoustic signal we extract features which are well known to capture word prominence like loudness, fundamental frequency and durational features. The analysis of the visual signal is based on features derived from the speaker’s rigid head movements and movements of the speaker’s mouth. We capture the rigid head movements by tracking the speaker’s nose. Via a two-dimensional Discrete Cosine Transform (DCT) calculated from the mouth region we represent the movements of the speaker’s mouth. The results show that the rigid head movements as well as movements inside the mouth region can be used to discriminate prominent from non-prominent words. The audio-only detection yields an Equal Error Rate (EER) averaged over all speakers of 13%. Based only on the visual features we obtain 20% of EER. When we combine the visual and the acoustic features we only see a small improvement compared to the audio-only detection for clean speech. To simulate background noise we added 4 different noise types at varying SNR levels to the acoustic stream. The results indicate that word prominence detection is quite robust against additional background noise. Even at a severe Signal to Noise Ratio (SNR) of − 10 dB the EER only rises to 35%. Despite this the audio-visual fusion leads to notable improvements for the detection from noisy speech. We observe relative reductions of the EER of up to 79%. [ABSTRACT FROM AUTHOR]

Precise and fast shape matching and retrieval from very large datasets is a challenging task because of the existence of many distortions such as noise, occlusion and affine distortions. In this paper, we aim to propose a time-saving and effective shape matching and retrieval framework, that employs pruning which will enable online shape retrieval from extremely large datasets. First, using a hierarchical tree-based structure supporting parallel processing and efficient feature descriptors, irrelevant shapes are pruned and a subset of shapes relevant to the query is selected, then using more sophisticated feature descriptors, accurate retrieval is performed. Contour representation of an object is considered as most significant visual shape similarity measure by the humans. Using boundary information, we generate two simplified and efficient feature descriptors for fast pruning, and a sophisticated feature descriptor for effective and accurate retrieval. Tests performed on standard datasets unveil that the proposed technique is computationally more efficient than the state-of-the-art techniques while maintaining comparable matching and retrieval performance. Its performance is scalable for huge datasets and is robust against affine transformations, articulations and occlusion. [ABSTRACT FROM AUTHOR]

Image features are widely used for object identification in many situations, including interpretation of data containing natural scenes captured by unmanned aerial vehicles. This paper presents a parallel framework to extract additive features (such as color features and histogram of oriented gradients) using the processing power of GPUs and multicore CPUs to accelerate the algorithms with the OpenCL language. The resulting features are available in device memory and then can be fed into classifiers such as SVM, logistic regression and boosting methods for object recognition. It is possible to extract multiple features with better performance. The GPU accelerated image integral algorithm speeds up computations up to 35x when compared to the single-thread CPU implementation in a test bed hardware. The proposed framework allows real-time extraction of a very large number of image features from full-HD images (better than 30 fps) and makes them available for access in coalesced order by GPU classification algorithms. [ABSTRACT FROM AUTHOR]

The paper describes a system for automatic speech recognition (ASR) that is benchmarked with data of the 3rd CHiME challenge, a dataset comprising distant microphone recordings of noisy acoustic scenes in public environments. The proposed ASR system employs various methods to increase recognition accuracy and noise robustness. Two different multi-channel speech enhancement techniques are used to eliminate interfering sounds in the audio stream. One speech enhancement method aims at separating the target speaker's voice from background sources based on non-negative matrix factorization (NMF) using variational Bayesian (VB) inference to estimate NMF parameters. The second technique is based on a time-varying minimum variance distortionless response (MVDR) beamformer that uses spatial information to suppress sound signals not arriving from a desired direction. Prior to speech enhancement, a microphone channel failure detector is applied that is based on cross-comparing channels using a modulation-spectral representation of the speech signal. ASR feature extraction employs the amplitude modulation filter bank (AMFB) that implicates prior information of speech to analyze its temporal dynamics. AMFBs outperform the commonly used frame splicing technique of filter bank features in conjunction with a deep neural network (DNN) based ASR system, which denotes an equivalent data-driven approach to extract modulation-spectral information. In addition, features are speaker adapted, a recurrent neural network (RNN) is employed for language modeling, and hypotheses of different ASR systems are combined to further enhance the recognition accuracy. The proposed ASR system achieves an absolute word error rate (WER) of 5.67% on the real evaluation test data, which is 0.16% lower compared to the best score reported within the 3rd CHiME challenge. [ABSTRACT FROM AUTHOR]

This paper gives an in-depth presentation of the multi-microphone speech recognition system we submitted to the 3rd CHiME speech separation and recognition challenge (CHiME-3) and its extension. The proposed system takes advantage of recurrent neural networks (RNNs) throughout the model from the front-end speech enhancement to the language modeling. Three different types of beamforming are used to combine multi-microphone signals to obtain a single higher-quality signal. The beamformed signal is further processed by a single-channel long short-term memory (LSTM) enhancement network, which is used to extract stacked mel-frequency cepstral coefficients (MFCC) features. In addition, the beamformed signal is processed by two proposed noise-robust feature extraction methods. All features are used for decoding in speech recognition systems with deep neural network (DNN) based acoustic models and large-scale RNN language models to achieve high recognition accuracy in noisy environments. Our training methodology includes multi-channel noisy data training and speaker adaptive training, whereas at test time model combination is used to improve generalization. Results on the CHiME-3 benchmark show that the full set of techniques substantially reduced the word error rate (WER). Combining hypotheses from different beamforming and robust-feature systems ultimately achieved 5.05% WER for the real-test data, an 84.7% reduction relative to the baseline of 32.99% WER and a 44.5% reduction from our official CHiME-3 challenge result of 9.1% WER. Furthermore, this final result is better than the best result (5.8% WER) reported in the CHiME-3 challenge. [ABSTRACT FROM AUTHOR]

This paper investigates the problem of measuring speech quality of received speech signal without employing the original speech signal. The problem of deterioration of the speech quality arises due to noise present in the surroundings. To this line, we propose a multiple time-instances (MTI) features-based approach for reference-free speech quality measurement model. A voice activity detector (VAD) is exploited primarily for calculating the number of active speech chunks of a speech signal. For these chunks and their successive combinations called here batches, multi-resolution auditory model (MRAM), mel-frequency cepstral coefficients (MFCC) and line spectral frequencies (LSF) features are extracted and called as MTI features. It is hypothesized that the MTI features are capable in capturing the distortions caused by time-localized effects of short-time transients, impulsive noise, and its differences from the plosive sounds. The MTI metric estimates (MTI-ME) are calculated corresponding to these MTI features employing the Gaussian mixture model (GMM) probabilistic technique. The overall objective speech quality of a speech signal is then determined as a linear combination of optimally weighted MTI-ME corresponding to distinct active speech chunks and their successive combinations, that is, batches of that speech signal. Minimum mean square error criterion or Pearson's correlation maximization criterion is employed for computing optimal weights. In addition, a deep neural network (DNN)-based speech quality model is also developed for calculating a single objective speech quality while considering all active speech chunks together. Pearson's correlation coefficient and weighted average correlation are exploited for evaluating the performance. Results demonstrate that the proposed model achieves promising improvement over the standard speech quality model (P.563) and improves correlation values by around 37%. • Addressing speech quality measuring problem by developing a reference-free speech quality model. • Developing a feature extraction technique incorporating distinct auditory features. • Developing a joint GMM training algorithm for computing objective speech quality. • Developing a deep neural network (DNN) framework for validating the proposed algorithm. • Demonstrating better performance as compared to the standard speech quality model (P.563). [ABSTRACT FROM AUTHOR]

This paper focuses on the task of cross-language relation extraction, which aims to identify the semantic relations holding between entities in the text. The goal of the task is to train classifiers for low-resource languages by means of the annotated data from high-resource languages. Related methods usually employ parallel data or Machine Translator (MT) to project annotated data from a source to a target language. However, the availability and the quality of parallel data and MT are big challenges for low-resource languages. In this paper, a novel transfer learning method is presented for this task. The key idea is to utilize a tree-based representation of data, which is highly informative for classifying semantic relations, and also shared among different languages. All the training and test data are shown using this representation. We propose to use the Universal Dependency (UD) parsing, which is a language-agnostic formalism for representation of syntactic structures. Equipping UD parse trees with multi-lingual word embeddings makes an ideal representation for the cross-language relation extraction task. We propose two deep networks to use this representation. The first one utilizes the Shortest Dependency Path of UD trees, while the second employs the UD-based positional embeddings. Experiments are performed using SemEval 2010-task 8 training data, whereas French and Farsi are the test languages. The results show 63.9% and 56.2% F 1 scores, for French and Farsi test data, respectively, which are 14.4% and 17.9% higher than the baseline. This work can be considered a simple yet powerful baseline for further investigation into the cross-language tasks. • A new method for cross-language relation extraction, which is suitable for low-resource languages. • Introduction of a new baseline for cross-language relation extraction. • Utilizing Universal Dependency parse formalism, for making a machine learning model to solve relation extraction problem. • This research opens up avenues for further research on cross-language problems. [ABSTRACT FROM AUTHOR]

Word Sense Disambiguation (WSD) is a fundamental task useful for Information Retrieval, Information Extraction, web search, and indexing, among others. In the literature there exist several works dedicated to generic WSD task, but in recent years domain-specific WSD has attracted the attention of several researchers. In this sense, this paper describes an approach for domain-specific WSD by selecting the predominant sense (synset from WordNet) of ambiguous words. To achieve it the method uses two corpora: the domain-specific test corpus (containing target ambiguous words) and a domain-specific auxiliary corpus (obtained by using relevant words from the domain-specific test corpus ). The approach has four main stages: (1) auxiliary corpus generation; (2) related features extraction (from the auxiliary corpus); (3) test features extraction (from the test corpus); and (4) features integration . The proposed approach has been tested on domain-specific corpora (Sports and Finance) and on one balanced corpus, BNC. Even though our WSD approach showed some limitations when dealing with the general-domain corpus, the obtained results for domain-specific corpora, which are our main interest, were better than those reported in previous works. [ABSTRACT FROM AUTHOR]

Bidirectional long short-term memory (Bi-LSTM), as one of the effective networks for sequence labeling tasks, is widely used in named entity recognition (NER). However, the sequential nature of Bi-LSTM and the inability to recognize multiple sentences at the same time make it impossible to obtain overall information. In this paper, to make up for the shortcomings of Bi-LSTM in extracting global information, we propose a hierarchical context model embedded with sentence-level and document-level feature extraction. In sentence-level feature extraction, we use the self-attention mechanism to extract sentence-level representations considering the different contribution of each word to the sentence. For document-level feature extraction, 3D convolutional neural network (CNN), which not only can extract features within sentences, but also pays attention to the sequential relationship between sentences, is used to extract document-level representations. Furthermore, we investigate a layer-by-layer residual (LBL Residual) structure to optimize each Bi-LSTM block of our model, which can solve the degradation problem that appears as the number of model layers increases. Experiments show that our model achieves results competitive with the state-of-the-art records on the CONLL-2003 and Ontonotes5.0 English datasets respectively. [ABSTRACT FROM AUTHOR]

The deep learning methodologies in state-of-the-art speaker recognition systems are predominantly limited to the extraction of recording level embeddings. This is usually followed by generative modeling of the embeddings to output the verification score. In this paper, we explore a fully neural approach where the neural model outputs the verification score directly, given the acoustic feature inputs. This model, termed as Siamese neural network (SiamNN), combines the embedding extraction and back-end modeling into a single processing pipeline. The back-end modeling is achieved using a neural approach to PLDA modeling, called neural probabilistic linear discriminant analysis (NPLDA). In the NPLDA model, the verification score is computed as a discriminative similarity function. The development of the single neural SiamNN model allows the joint optimization of all the modules using a verification cost. Several speaker recognition experiments are performed using SITW, VOiCES, and NIST SRE datasets where the proposed SiamNN model is shown to significantly improve over the state-of-art x-vector PLDA baseline system (relative improvements of up to 35% in the primary cost metric). We also provide a detailed analysis of the influence of hyper-parameters, choice of loss functions, and data sampling strategies for training the model. In particular, we highlight that the proposed soft detection cost function based optimization improves over other loss functions considered. [ABSTRACT FROM AUTHOR]

Currently, multimedia data has become one of the most important data types processed and transferred over the Internet. To extract useful information from a huge amount of such data, SIFT and SURF, as two most popular image feature extraction algorithms, have been widely used in many applications running on multi-core platforms. However, limited parallelism in existing designs makes it hard or impossible to apply them in many applications with real-time requirements. Therefore, it has become one of the major challenges to improve the processing speed of image feature extraction algorithms. In this paper, we first analyze the parallelism constraints in the algorithms, such as imbalanced workloads and indeterminate time distributions. Based on such analyses, we present an adaptive pipeline parallel scheme (AD-PIPE) to adjust the thread number in different stages according to their workloads dynamically, which achieves a balanced partition for constant input workloads. Furthermore, we also implement a power efficient version (AE-PIPE) for AD-PIPE through scheduling threads based on variable input workloads. Experimental results show that AD-PIPE achieves a speedup of 16.88X and 20.33X respectively over SIFT and SURF on a 16-core machine. Moreover, AE-PIPE achieves up to 52.94% and 58.82% power saving with only 3% performance loss. [ABSTRACT FROM AUTHOR]

In this paper we present Ariadne , a compiler that extracts parallelism from recursive function calls. Ariadne takes as input C code enhanced with directives for recursive functions and automatically produces code for multi-core architectures. It produces code for the POSIX standard, the OpenMP model and the Cilk programming language, which run on a wide variety of computing systems. Ariadne also produces code for SL, a programming language proposed for the SVP processor and model. This is of special interest, since we can map certain function calls onto SVP, which contain inherent parallelism that cannot efficiently be expressed in other programming models. Ariadne is the only compiler that extracts parallelism from various forms of recursive functions using directives. It is also the only compiler that handles all forms of reduction operations for addition, subtraction, multiplication and division. The experimental results are very promising showing significant speedups in all benchmarks. [ABSTRACT FROM AUTHOR]

The term assortativity indicates the tendency, for a network node, to be directly connected to other nodes that are someway similar. In more technical terms, a given feature is assortative in a network if the probability that an arc exists between two nodes having this feature is greater than the probability that an arc exists between two generic nodes. The role of assortativity in real-world and online social networks has been largely investigated in the literature, in which, starting from degree assortativity, several forms of assortativity have been analyzed. When moving from a single-social-network to a multiple-social-network perspective, new specific traits can be studied, also under the assortativity magnifying glass. This is the case of membership overlap among networks (i.e., the fact that people belong to more online social networks) as expression of different traits of users׳ personality. In this paper, we deal with the above issue, by defining two different measures of membership overlap assortativity, called Loose and Constrained Inter-social-network Assortativity , respectively and by observing that in two of the most representative online social networks, namely Facebook and Twitter , membership overlap is assortative. [ABSTRACT FROM AUTHOR]

• Dance teachers face challenges to provide learners with feedback on rhythmic skills. • Current rhythm modelling solutions are too complex to support dance teachers 'in-thewild'. • We present an approach to model learners' rhythm, using a single smartphone, to enrich teachers' assessments. • Resulting models were validated by comparing the artificial intelligence (AI) outputs with dance teachers' assessments. • We investigated teachers' envisaged pedagogical uses of the AI to support rhythm learning. Rhythmic skills are fundamental in social partner dance. Yet, large class sizes and time constraints often make it challenging for teachers to assess learners' rhythmic skills to provide constructive feedback. As motion sensors are becoming widely available and embedded in most smartphones, it is also becoming feasible to create models of anatomical movements of the human body that could be used to support such teachers' assessments. Current solutions aimed at modelling motor skills using motion sensors and artificial intelligence (AI) have enabled the detection of some dimensions of rhythm, but they are either technically too complex to be used 'in-the-wild' or have not been designed based on dance teachers' assessment needs. This paper presents an approach for modelling rhythmic movements, via a single smartphone, while learners learn how to dance Forró. The purpose is to enrich teachers' rhythmic assessments with data. Following a user-centred approach, we first elicited from teachers the dance elements that they commonly focus their rhythm assessments on (i.e., tempo, pause, step size and weight transfer). Then, selected features were extracted from raw motion sensor data related to the rhythmic patterns of learners dancing, their synchronisation with the beat of the music, and particular characteristics of the song being played. Finally, machine learning (ML) algorithms were used to create predictive computational models using these features. The modelling approach was validated through two studies: (1) a quantitative comparison between the ML outputs and dance teachers assessments of learners' dance performance; and (2) a qualitative analysis of the potential pedagogical uses of the outputs of the ML models envisioned by dance teachers. [ABSTRACT FROM AUTHOR]

Random Indexing based extractive text summarization has already been proposed in literature. This paper looks at the above technique in detail, and proposes several improvements. The improvements are both in terms of formation of index (word) vectors of the document, and construction of context vectors by using convolution instead of addition operation on the index vectors. Experiments have been conducted using both angular and linear distances as metrics for proximity. As a consequence, three improved versions of the algorithm, viz. RISUM, RISUM+ and MRISUM were obtained. These algorithms have been applied on DUC 2002 documents, and their comparative performance has been studied. Different ROUGE metrics have been used for performance evaluation. While RISUM and RISUM+ perform almost at par, MRISUM is found to outperform both RISUM and RISUM+ significantly. MRISUM also outperforms LSA+TRM based summarization approach. The study reveals that all the three Random Indexing based techniques proposed in this study produce consistent results when linear distance is used for measuring proximity. [ABSTRACT FROM AUTHOR]

Recently, re-ranking algorithms have been successfully applied on statistical machine translation systems. Due to the errors in the hypothesis alignment and varying word order between the source and target sentences and also the lack of sufficient resources such as parallel corpora, decoding may result in ungrammatical or non-fluent outputs. This paper proposes a re-ranking system based on swarm algorithms, which makes the use of sophisticated non-syntactical features to re-rank the n-best translation candidates. We introduce plenty of easy-computed non-syntactical features to deal with SMT system errors plus the quantum-behaved particle swarm optimization (QPSO) algorithm to adjust the weights of features. We have evaluated the proposed approach on 2 translation tasks in different language pairs (Persian → English and German → English) and genres (news and novel books). In comparison with PSO-, GA-, Perceptron- and averaged Perceptron-style re-ranking systems, the experimental study demonstrates the superiority of the proposed system in terms of translation quality on both translation tasks. In addition, the impacts of the proposed features on the translation quality have been analyzed, and the most positive ones have been recognized. At the end, the impact of the n-best list size on the proposed system is investigated. [ABSTRACT FROM AUTHOR]

Empirical Mode Decomposition (EMD) was designed to analyze nonlinear and non-stationary signals. EMD voice analysis had been applied to Parkinson's sustained vowels, but very limited studies have been done on highly dynamic Diadochokinesia (DDK) utterances. This paper applies the EMD's dyadic filterbank characteristics to extract DDK features and an in-depth study on the efficacy of two segmentation strategies. The EMD analysis on DDK looks at the spectrum characteristics of Intrinsic Mode Functions (IMF) and the handling of mode mixing conditions. DDK recordings of Healthy Control (HC) subjects and patients with Parkinson's disease (PD) were segmented using various fixed frame sizes compared to dynamic segmentation based on/pa-ta-ka/ triad length, and also the signal envelope as a whole. An overlapping windowing of 2/3 was used in the fixed frame size segmentation to augment and to capture the redundant and transition information. No overlapping was used in the/pa-ta-ka/ triad segmentation. For the fixed frame size segmentation, we found that there is a region of consistency. Within this region, the IMF center frequencies and bandwidths maintained the same but varied outside the region. The segmentation comparisons used a basic set of EMD features with and without DeltaEMD features that capture segment-to-segment deviations. Using the basic EMD dyadic features, fixed frame size segmentation out-performed/pa-ta-ka/ triad segmentation. When DeltaEMD features were added to provide segment deviation information,/pa-ta-ka/ triad out-performed fixed frame segmentation. Additional segment-magnitude amplification factor and segment length were found to improve the performance of the/pa-ta-ka/ triad segmentation. With the added features,/pa-ta-ka/ triad out-performed the others and had an improved accuracy of 78%. Additional features have also increased the envelope discrimination to 76%. The results also indicated the potentials of using voice envelopes for PD analysis. [Display omitted] • Parkinson's Diadochokinesia/pa-ta-ka/ utterance is studied. • EMD dyadic characteristic is applied to extract articulation of DDK and dynamic of the signal envelopes. • Proposed EMD and DeltaEMD dyadic features to capture the articulation of nonlinear and non-stationary DDK utterance. • Fixed frame size and/pa-ta-ka/ triad segmentation strategies are compared. [ABSTRACT FROM AUTHOR]

• The vowel transition regions lie in the junction between a consonant and vowel (CV) regions and between a vowel and a consonant (VC) regions. • We have proposed a method for determining the vowel transition regions based on the rate of change of formant frequencies using zero-time windowing and numerator of group-delay function. • Zero-time windowing derives the instantaneous formant frequencies accurately at every sample location due to contribution of that sample itself. • The numerator of group-delay function enhances the formant frequencies. • The proposed transition region detection method is evaluated on CV, CVC and TIMIT databases and it has shown significant improvement in the performance compared to the existing method. The vowel transition regions are the crucial landmarks in the speech signal. These vital regions are present at both ends of the vowel. They lie in the junction between a consonant and a vowel (CV) regions. This region plays an important role in numerous speech applications like speaker recognition, emotion conversion, speech rate modification, and CV unit recognition. The performance of these applications crucially depends on the accuracy of the estimation of vowel transition regions. In this paper, we have proposed a method for determining the transition regions based on the rate of change of formant frequencies using zero-time windowing and numerator of the group-delay function. Zero-time windowing derives the instantaneous formant frequencies accurately at every sample location due to the contribution of that sample itself. The numerator of the group-delay function enhances the formant frequencies. The proposed transition region detection method is evaluated on CV, and continuous speech databases recorded in the Hindi language. The proposed method has shown around 12% improvement in accuracy compared to the existing method. [ABSTRACT FROM AUTHOR]

• Propose a novel self-segmentation of pass-phrase utterances for deep feature extraction. • Introduce gender-dependent, speaker-adapted phone HMMs for segmentation and label. • Eliminate the need of a general-purpose, potentially-mismatched ASR for segmentation. • Propose a bottleneck feature trained to discriminate gender-dependent phones. • Study fusion of the proposed and existing features in score and feature domains. In this paper, we propose a novel method to segment and label pass-phrase utterances for training deep neural network (DNN) bottleneck (BN) features for text-dependent speaker verification (TD-SV). Specifically, gender-dependent hidden Markov models (HMMs) for monophones are first trained using the pass-phrase utterances that are disjoint from evaluation. Next, the trained HMMs are speaker-adapted and then used for segmenting and labeling these training utterances at the phone level. The resulted labeled data is subsequently used for training DNN models to discriminate gender-dependent phones for the purpose of extracting phone-discriminant BN features. This is in contrast to conventional approaches that apply a general-purpose, speaker-independent automatic speech recognition (ASR) system for generating segmentation and labels. The proposed method eliminates the need for a separate ASR system, which can additionally have the disadvantage of mismatch with the pass-phrase utterances in terms languages, dialects, domains, acoustic conditions and so on. Experiments are conducted on the RedDots challenge 2016 database of TD-SV using short utterances with Gaussian mixture model-universal background model and i-vector techniques. Experimental results demonstrate that the proposed method yields lower error rates in TD-SV when compared to a set of existing methods. A thorough ablation study further confirms the effectiveness of the method. Fusion in both score and feature levels also shows the complementary nature of the proposed features. [ABSTRACT FROM AUTHOR]

• Chroma based musical features are explored to capture rhythmic and music related features among dialects of a Kannada a Dravidian language. • Spectral statistical characteristics are computed from short term spectra in terms of spectral shape features. • A combination of Chroma and Spectral shape features is evaluated on the Kannada and English dialects. • The single classifier based support vector machine (SVM) and ensemble based SVM algorithms are employed for evaluation of the derived combination of features. • The system is found to be more suitable for forensic applications for speaker profiling tasks. • The proposed method has shown sustained improvement over the baseline system under limited data and noisy environment. The present work proposes a text-independent dialect identification system. Generally, dialects of a language exhibit varying pronunciation styles followed in a specific geographical region. In this paper, chroma features familiar with music-related systems are employed for identification of dialects. In addition, eight significant spectral shape related features from short term spectra are computed and combined along with chroma features and named as chroma-spectral shape features. Chroma features try to aggregate spectral information and attempt to encapsulate the evidential variations, concerning timbre, correlated melody, rhythmic, and intonation patterns found prominently among dialects of few languages. The effectiveness of the proposed features and approach is evaluated on five prominent Kannada dialects spoken in Karnataka, India and internationally known standard Intonation Variation in English (IViE) dataset with nine British English dialects. Discriminative models such as, single classifier based Support Vector Machine (SVM) and ensemble based support vector machines (ESVM) are employed for classification. The proposed features have shown better performance over state-of-the-art i-vector features on both datasets. The highest recognition performance of 95.6% and 97.52% are achieved in the cases of Kannada and IViE dialect datasets respectively using ESVM. Proposed features have also demonstrated robust performance with small sized (limited data) audio clips even in noisy conditions. [ABSTRACT FROM AUTHOR]

• A novel excitation model based on excitation source parameters and phone-specific natural residual segments is proposed for accurate parameterization and generation of excitation signals in SPSS framework. • Energy, Source spectrum, epoch strength and sharpness are considered as the excitation features. • The effectiveness of proposed excitation model is analyzed in HMM-based and DNN-based Statistical parameteric speech synthesis frameworks. In this paper, a novel excitation modelling method is proposed for improving the naturalness of statistical parametric speech synthesis (SPSS). In the proposed approach, the excitation or residual signal is parameterized by using features extracted from the epochs. The epoch parameters used in this work are epoch strength and sharpness. These features are modeled in the statistical framework along with other parameters. During synthesis, the excitation signal is constructed by imposing the generated epoch parameters on the natural instances of excitation signal. The effectiveness of the proposed method is evaluated in the framework of hidden Markov model (HMM)-based and deep neural network (DNN)-based SPSS. Evaluation results have shown that the SPSS systems developed using the proposed excitation model are capable of synthesizing more natural sounding speech compared to the ones based on two state-of-the-art excitation modelling approaches. [ABSTRACT FROM AUTHOR]