Feature Transformation Techniques Research Articles

A Framework for Detecting Thyroid Cancer from Ultrasound and Histopathological Images Using Deep Learning, Meta-Heuristics, and MCDM Algorithms.

Computer-assisted diagnostic systems have been developed to aid doctors in diagnosing thyroid-related abnormalities. The aim of this research is to improve the diagnosis accuracy of thyroid abnormality detection models that can be utilized to alleviate undue pressure on healthcare professionals. In this research, we proposed deep learning, metaheuristics, and a MCDM algorithms-based framework to detect thyroid-related abnormalities from ultrasound and histopathological images. The proposed method uses three recently developed deep learning techniques (DeiT, Swin Transformer, and Mixer-MLP) to extract features from the thyroid image datasets. The feature extraction techniques are based on the Image Transformer and MLP models. There is a large number of redundant features that can overfit the classifiers and reduce the generalization capabilities of the classifiers. In order to avoid the overfitting problem, six feature transformation techniques (PCA, TSVD, FastICA, ISOMAP, LLE, and UMP) are analyzed to reduce the dimensionality of the data. There are five different classifiers (LR, NB, SVC, KNN, and RF) evaluated using the 5-fold stratified cross-validation technique on the transformed dataset. Both datasets exhibit large class imbalances and hence, the stratified cross-validation technique is used to evaluate the performance. The MEREC-TOPSIS MCDM technique is used for ranking the evaluated models at different analysis stages. In the first stage, the best feature extraction and classification techniques are chosen, whereas, in the second stage, the best dimensionality reduction method is evaluated in wrapper feature selection mode. Two best-ranked models are further selected for the weighted average ensemble learning and features selection using the recently proposed meta-heuristics FOX-optimization algorithm. The PCA+FOX optimization-based feature selection + random forest model achieved the highest TOPSIS score and performed exceptionally well with an accuracy of 99.13%, F2-score of 98.82%, and AUC-ROC score of 99.13% on the ultrasound dataset. Similarly, the model achieved an accuracy score of 90.65%, an F2-score of 92.01%, and an AUC-ROC score of 95.48% on the histopathological dataset. This study exploits the combination novelty of different algorithms in order to improve the thyroid cancer diagnosis capabilities. This proposed framework outperforms the current state-of-the-art diagnostic methods for thyroid-related abnormalities in ultrasound and histopathological datasets and can significantly aid medical professionals by reducing the excessive burden on the medical fraternity.

Modelling bank customer behaviour using feature engineering and classification techniques

This study investigates customer behaviour and activity in the banking sector and uses various feature transformation techniques to convert the behavioural data into different data structures. Feature selection is then performed to generate feature subsets from the transformed datasets. Several classification methods used in the literature are applied to the original and transformed feature subsets. The proposed combined knowledge mining model enable us to conduct a benchmark study on the prediction of bank customer behaviour. A real bank customer dataset, drawn from 24,000 active and inactive customers, is used for an experimental analysis, which sheds new light on the role of feature engineering in bank customer classification. This paper’s detailed systematic analysis of the modelling of bank customer behaviour can help banking institutions take the right steps to increase their customers’ activity.

Classification of the perceptual impression of source-level blending between violins in a joint performance

Quantifying auditory perception of blending between sound sources is a relevant topic in music perception, but remains poorly explored due to its complex and multidimensional nature. Previous studies were able to explain the source-level blending in musically constrained sound samples, but comprehensive modelling of blending perception that involves musically realistic samples was beyond their scope. Combining the methods of Music Information Retrieval (MIR) and Machine Learning (ML), this investigation attempts to classify sound samples from real musical scenarios having different musical excerpts according to their overall source-level blending impression. Monophonically rendered samples of 2 violins in unison, extracted from in-situ close-mic recordings of ensemble performance, were perceptually evaluated and labeled into blended and non-blended classes by a group of expert listeners. Mel Frequency Cepstral Coefficients (MFCCs) were extracted, and a classification model was developed using linear and non-linear feature transformation techniques adapted from the dimensionality reduction strategies such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and t-Stochastic Neighbourhood Embedding (t-SNE), paired with Euclidean distance measure as a metric to evaluate the similarity of transformed feature clusters. Results showed that LDA transformed raw MFCCs trained and validated using a separate train-test data set and Leave-One-Out Cross-Validation (LOOCV) resulted in an accuracy of 87.5%, and 87.1% respectively in correctly classifying the samples into blended and non-blended classes. In this regard, the proposed classification model which incorporates “ecological” score-independent sound samples without requiring access to individual source recordings advances the holistic modeling of blending.

New Pharmacophore Fingerprints and Weight-matrix Learning for Virtual Screening. Application to Bcr-Abl Data.

In this work, we propose to analyze the potential of a new type of pharmacophoric descriptors coupled to a novel feature transformation technique, called Weight-Matrix Learning (WML, based on a feed-forward neural network). The application concerns virtual screening on a tyrosine kinase named BCR-ABL. First, the compounds were described using three different families of descriptors: our new pharmacophoric descriptors, and two circular fingerprints, ECFP4 and FCFP4. Afterwards, each of these original molecular representations were transformed using either an unsupervised WML method or a supervised one. Finally, using these transformed representations, K-Means clustering algorithm was applied to automatically partition the molecules. Combining our pharmacophoric descriptors with supervised Weight-Matrix Learning (SWMLR ) leads to clearly superior results in terms of several quality measures.

Predicting the prevalence of lung cancer using feature transformation techniques

Healthcare sector is one of the most important sectors of any country as a big part of the country’s economy is associated with it. The research is about to contribute to the health sector by minimizing the expenses of a lung cancer diagnosis. The study tries to devise an efficient method for initial screening of the patients with symptoms through their demographic and clinical data. The study seeks appropriate feature transformation techniques from dimensionality reduction techniques in combination with an apposite regression model that can perform this task robustly using the lung cancer dataset for early carcinoma diagnosis. To equip the health sector with state-of-the-art technology and for the betterment of humankind, the most beneficial tool of today is machine learning. Lungs play a vivacious role in the human body, oxygen is circulated in the body, and the air is taken from the atmosphere and sends it into the bloodstream. Several people decease every year because of lung carcinoma as lung cancer is normally identified at the latter phase due to lack of awareness. It stays unidentified because people have pneumonia often, which converts into lung cancer later. The projected research seeks to enable health professionals to the rationalization of primary diagnosis and treatment of lung carcinoma in developing countries. The proposed methodology has selected the optimized combination of regression-based machine learning technique and feature transformation technique for the available patterns based on demographic and clinical features of lung cancer patients. Based on the gathered results during training (RMSE = 0.1324, R2-Score = 0.7428) and testing (RMSE = 0.1273, R2-Score = 0.7405) it can be concluded that the fast independent component analysis and elastic net regression technique provided optimized results and outperform the other aspirant techniques.

A Review on Speech Emotion Recognition Using Machine Learning

This paper focuses on the development of a robust speech emotion recognition system using a combination of different speech features with feature optimization techniques and speech de-noising technique to acquire improved emotion classification accuracy, decreasing the system complexity and obtain noise robustness. Additionally, we create original methods for SER to merge features. We employ feature optimization methods that are based on the feature transformation and feature selection machine learning techniques in order to build SER. The following is a list of the upcoming events. A neural network can use either of these two techniques. As more feelings are taken into account, the feature fusion-acquired SER accuracy falls short of expectations, and the plague of dimensionality starts to spread due to the addition of speech features, which makes the SER system work harder to complete its task. This is due to the SER system becoming more complicated when voice elements are added. Therefore, it is crucial to create a SER system that is more trustworthy, has the most practical features, and uses the least amount of computing power possible. By using strategies that maximize current features, it is possible to streamline the feature selection process by reducing the total number of accessible choices to a more reasonable level. This piece employs a method known as Semi-Non Negative Matrix Factorization to lessen the amount of processing trash that the SER system generates. (Semi-NMF). This approach can be used to change traits that are capable of learning on their own.

A MACHINE LEARNING CLASSIFICATION APPROACH TO DETECT TLS-BASED MALWARE USING ENTROPY-BASED FLOW SET FEATURES

Transport Layer Security (TLS) based malware is one of the most hazardous malware types, as it relies on encryption to conceal connections. Due to the complexity of TLS traffic decryption, several anomaly-based detection studies have been conducted to detect TLS-based malware using different features and machine learning (ML) algorithms. However, most of these studies utilized flow features with no feature transformation or relied on inefficient flow feature transformations like frequency-based periodicity analysis and outliers percentage. This paper introduces TLSMalDetect, a TLS-based malware detection approach that integrates periodicity-independent entropy-based flow set (EFS) features generated by a flow feature transformation technique to solve flow feature utilization issues in related research. EFS features effectiveness was evaluated in two ways: (1) by comparing them to the corresponding outliers percentage and flow features using four feature importance methods, and (2) by analyzing classification performance with and without EFS features. Moreover, new Transmission Control Protocol features not explored in literature were incorporated into TLSMalDetect, and their contribution was assessed. This study’s results proved EFS features of the number of packets sent and received were superior to related outliers percentage and flow features and could remarkably increase the performance up to ~42% in the case of Support Vector Machine accuracy. Furthermore, using the basic features, TLSMalDetect achieved the highest accuracy of 93.69% by Naïve Bayes (NB) among the ML algorithms applied. Also, from a comparison view, TLSMalDetect’s Random Forest precision of 98.99% and NB recall of 92.91% exceeded the best relevant findings of previous studies. These comparative results demonstrated the TLSMalDetect’s ability to detect more malware flows out of total malicious flows than existing works. It could also generate more actual alerts from overall alerts than earlier research.Transport Layer Security (TLS) based malware is one of the most hazardous malware types, as it relies on encryption to conceal connections. Due to the complexity of TLS traffic decryption, several anomaly-based detection studies have been conducted to detect TLS-based malware using different features and machine learning (ML) algorithms. However, most of these studies utilized flow features with no feature transformation or relied on inefficient flow feature transformations like frequency-based periodicity analysis and outliers percentage. This paper introduces TLSMalDetect, a TLS-based malware detection approach that integrates periodicity-independent entropy-based flow set (EFS) features generated by a flow feature transformation technique to solve flow feature utilization issues in related research. EFS features effectiveness was evaluated in two ways: (1) by comparing them to the corresponding outliers percentage and flow features using four feature importance methods, and (2) by analyzing classification performance with and without EFS features. Moreover, new Transmission Control Protocol features not explored in literature were incorporated into TLSMalDetect, and their contribution was assessed. This study’s results proved EFS features of the number of packets sent and received were superior to related outliers percentage and flow features and could remarkably increase the performance up to ~42% in the case of Support Vector Machine accuracy. Furthermore, using the basic features, TLSMalDetect achieved the highest accuracy of 93.69% by Naïve Bayes (NB) among the ML algorithms applied. Also, from a comparison view, TLSMalDetect’s Random Forest precision of 98.99% and NB recall of 92.91% exceeded the best relevant findings of previous studies. These comparative results demonstrated the TLSMalDetect’s ability to detect more malware flows out of total malicious flows than existing works. It could also generate more actual alerts from overall alerts than earlier research.

Voltammetric Electrode Array Optimization for Black Tea Discrimination Using Computational Intelligence Approach

In this paper, a computational intelligence approach has been adopted to optimize the voltammetric electrode array for discrimination of the overall quality of black tea. Electrochemical measurements have been performed using the three- electrode system that comprises of an Ag/AgCl reference electrode, a platinum counter electrode, and synthesized working electrodes (WE). Nine working electrodes (WE1 to WE9) have been synthesized by varying the compositions of polymer and graphite. The electrodes have been imbibed in black tea samples and subjected to triangular voltage waveform ranging from -0.04 V to 0.8 V. From the data set so obtained, the number of electrodes have been optimized using computational approaches followed by a polling technique. Features were extracted by four feature transformation techniques and samples were classified with five different classification methods. Polling score has been assigned to each WE based on the decisions obtained by different feature transformation and classification techniques. The average classification accuracy rate of two working electrodes (WE1 and WE8) optimized from the algorithms were 91.27% and 91.34%, respectively. Thus, the optimization technique implemented in the present work yielded acceptable results and this technique could be suitable for other real time applications as well.

Investigating and improving the utility of probabilistic linear discriminant analysis for acoustic signal classification

Probabilistic linear discriminant analysis (PLDA) has achieved good performance in face recognition and speaker recognition. However, the computation of PLDA using the original formulation is inefficient when there are many training data, especially when the dimensionality of the data is high. Faced with this inefficiency issue, we propose scalable formulations for PLDA. The computation of PLDA using the scalable formulations is more efficient than using the original formulation when dealing with many training data. Using the scalable formulations, the PLDA model can significantly outperform other popular classifiers for speaker recognition, such as Support Vector Machine (SVM) and Gaussian Mixture Model (GMM). Besides of directly using PLDA as a classifier, we may also use PLDA as a feature transformation technique. This PLDA-based feature transformation technique can reduce or expand the original feature dimensionality, and at the same time keep the transformed feature vector approximately following the Gaussian distribution. Our experimental results on speaker recognition and acoustic scene classification demonstrate the effectiveness of applying PLDA for feature transformation. It is then promising to combine PLDA with other classification models for improved performance, extending the utility of PLDA to a wider range of areas.

Tax Default Prediction Using Feature Transformation-Based Machine Learning

This study proposes to address the economic significance of unpaid taxes by using an automatic system for predicting a tax default. Too little attention has been paid to tax default prediction in the past. Moreover, existing approaches tend to apply conventional statistical methods rather than advanced data analytic approaches, including state-of-the-art machine learning methods. Therefore, existing studies cannot effectively detect tax default information in real-world financial data because they fail to take into account the appropriate data transformations and nonlinear relationships between early-warning financial indicators and tax default behavior. To overcome these problems, this study applies diverse feature transformation techniques and state-of-the-art machine learning approaches. The proposed prediction system is validated by using a dataset showing tax defaults and non-defaults at Finnish limited liability firms. Our findings provide evidence for a major role of feature transformation, such as logarithmic and square-root transformation, in improving the performance of tax default prediction. We also show that extreme gradient boosting and the systematically developed forest of multiple decision trees outperform other machine learning methods in terms of accuracy and other classification performance measures. We show that the equity ratio, liquidity ratio, and debt-to-sales ratio are the most important indicators of tax defaults for 1-year-ahead predictions. Therefore, this study highlights the essential role of well-designed tax default prediction systems, which require a combination of feature transformation and machine learning methods. The effective implementation of an automatic tax default prediction system has important implications for tax administration and can assist administrators in achieving feasible government expenditure allocations and revenue expansions.

An effective intrusion detection approach using SVM with naïve Bayes feature embedding

Network security has become increasingly important in recent decades, while intrusion detection system plays a critical role in protecting it. Various machine learning techniques have been applied to intrusion detection, among which SVM has been considered as an effective method. However, existing studies rarely take the data quality into consideration, which is essential for constructing a well-performed intrusion detection system beyond machine learning techniques. In this paper, we propose an effective intrusion detection framework based on SVM with naïve Bayes feature embedding. Specifically, the naïve Bayes feature transformation technique is implemented on the original features to generate new data with high quality; then, an SVM classifier is trained using the transformed data to build the intrusion detection model. Experiments on multiple datasets in intrusion detection domain validate that the proposed detection method can achieve good and robust performances, with 93.75% accuracy on UNSW-NB15 dataset, 98.92% accuracy on CICIDS2017 dataset, 99.35% accuracy on NSL-KDD dataset and 98.58% accuracy on Kyoto 2006+ dataset. Furthermore, our method possesses huge advantages in terms of accuracy, detection rate and false alarm rate when compared to other methods.

Scalable fault models for diagnosis in a synchronous generator using feature mapping and transformation techniques

Condition based maintenance (CBM) needs data acquired during healthy and faulty conditions to develop intelligent system for fault diagnosis. However, fault injection is not allowed/possible in a highly expensive components of complex/critical systems to collect fault condition data. Therefore, proto-type/small working models are used to conduct experiments for abnormal/fault conditions, to obtain and scale the intelligence of the system for effective health monitoring of complex system. This methodology is referred as scalable fault models. For proof of concept, in this work, we considered two different capacity synchronous generators with rating of 3 kVA and 5 kVA to emulate the behavior of prototype/small working model and complex system respectively, for scalable fault models. We explored feature mapping and transformation techniques to achieve effective scalability.From the preliminary experiments, it is observed that the baseline system performance deteriorated due to the changes in the system (capacity) and its characteristics with load changes.We therefore, expressed the input features in terms of load and system independent manner, to make the features less dependent on load and system variations. We explored localityconstrained linear coding (LLC) to express the features load/system independently. It is observed that experimenting LLC with the backend support vector machine (SVM) classifier gave the best fault classification performance for linear kernel, suggesting that the faults are linearly separable in the new feature space.Since the LLC mapped feature space is linearly separable, we then explored linear feature transformation technique, nuisance attribute projection (NAP) on the LLC mapped feature space to further minimize the load/system specific variations. We observed that LLC-NAP improved the overall accuracy and sensitivity of the classifier significantly. We also noted that the performance of NAP was limited in the original feature space since the feature space (NAP without LLC) is nonlinear with load/system variations.

Similarity Based Feature Transformation for Network Anomaly Detection

The fundamental objective behind any network intrusion detection system is to automate the detection process whenever intrusions occur in the network. The problem of the network anomaly detection is to determine, if the network incoming traffic is legitimate (or) anomalous. Automated detection systems designed to identify incoming anomalous traffic patterns usually apply widely used machine learning techniques. However, irrespective of any system model which is developed to identify anomalous traffic, all these models requires comparing anomalous and normal traffic patterns. Such comparisons implicitly depend on the ability of the underlying machine learning model to gauge the similarity between a known legitimate observation and the target. The efficiency of any network anomalous detection system depends on the use of distance (or) similarity measures and how they are actually applied. A novel distance function which can be applied to determine the similarity between two conditional feature pattern vectors is an important contribution of present research. Feature dimensionality is another important issue for any machine learning algorithm. In the present work, feature reduction is achieved using the proposed feature transformation technique. However, our approach for feature transformation uses the proposed gaussian distance function to achieve dimensionality reduction to represent the original input dataset in the new transformation space. We have also proposed new computation expressions for determining equivalent deviation and threshold in gaussian space. Experiments are performed on KDD and NSL-KDD datasets by considering widely applied classifier algorithms in various state-of-art research contributions. For performance validation of machine learning models, k-fold cross validation is applied by setting k to 10 through considering evaluation parameters such as accuracy, precision and recall. Experiment results have proved that our approach for anomaly detection that applies the proposed feature transformation technique proved comparatively better to detection methods CANN, GARUDA, and UTTAMA addressed in the recent research literature.

Prediction of stock movement using phase space reconstruction and extreme learning machines

ABSTRACTStock movement prediction is regarded as one of the most difficult, meaningful, and attractive research issues in the field of financial markets. The stock price data have non-stationary, noisy, and non-linear characteristics which make the movement and its prediction a challenging task. In this paper, we propose a framework to predict the stock price movement using phase space reconstruction (PSR) and extreme learning machines (ELM). The uniqueness of the framework is reflected by its feature transformation technique which computes the information distance from the transformed features in phase space. The distance from phase space dimensions are modelled with ELM to predict the stock price movement. A decision-level fusion is performed on the ELM models trained using each category of features to improve the prediction performance. The framework has been validated on one of the challenging Borsa Istanbul (BIST 100) dataset which is a widely used dataset in stock price prediction studies. The results from the proposed framework are compared with the conventional machine learning pipeline as well as the baseline methods, i.e., random and Naïve approach to show the effectiveness in prediction performance. Experimental results reveal that the framework improves predictive performance by 4.5% in terms of F-measure values.

Performance tuning for machine learning-based software development effort prediction models

Software development effort estimation is a critical activity of the project management process. In this study, machine learning algorithms were investigated in conjunction with feature transformation, feature selection, and parameter tuning techniques to estimate the development effort accurately and a new model was proposed as part of an expert system. We preferred the most general-purpose algorithms, applied parameter optimization technique (Grid- Search), feature transformation techniques (binning and one-hot-encoding), and feature selection algorithm (principal component analysis). All the models were trained on the ISBSG datasets and implemented by using the scikit-learn package in the Python language. The proposed model uses a multilayer perceptron as its underlying algorithm, applies binning of the features to transform continuous features and one-hot-encoding technique to transform categorical data into numerical values as feature transformation techniques, does feature selection based on the principal component analysis method, and performs parameter tuning based on the GridSearch algorithm. We demonstrate that our effort prediction model mostly outperforms the other existing models in terms of prediction accuracy based on the mean absolute residual parameter.

Biometric template security: an overview

PurposeThe security of the stored biometric template is itself a challenge. Feature transformation techniques and biometric cryptosystems are used to address the concerns and improve the general acceptance of biometrics. The purpose of this paper is to provide an overview of different techniques and processes for securing the biometric templates. Furthermore, the paper explores current research trends in this area.Design/methodology/approachIn this paper, the authors provide an overview and survey of different features transformation techniques and biometric cryptosystems.FindingsFeature transformation techniques and biometric cryptosystems provide reliable biometric security at a high level. There are many techniques that provide provable security with practical viable recognition rates. However, there remain several issues and challenges that are being faced during the deployment of these technologies.Originality/valueThis paper provides an overview of currently used techniques for securing biometric templates and also outlines the related issues and challenges.

Evaluating Interactive Visualization of Multidimensional Data Projection with Feature Transformation

There has been extensive research on dimensionality reduction techniques. While these make it possible to present visually the high-dimensional data in 2D or 3D, it remains a challenge for users to make sense of such projected data. Recently, interactive techniques, such as Feature Transformation, have been introduced to address this. This paper describes a user study that was designed to understand how the feature transformation techniques affect user’s understanding of multi-dimensional data visualisation. It was compared with the traditional dimension reduction techniques, both unsupervised (PCA) and supervised (MCML). Thirty-one participants were recruited to detect visual clusters and outliers using visualisations produced by these techniques. Six different datasets with a range of dimensionality and data size were used in the experiment. Five of these are benchmark datasets, which makes it possible to compare with other studies using the same datasets. Both task accuracy and completion time were recorded for comparison. The results show that there is a strong case for the feature transformation technique. Participants performed best with the visualisations produced with high-level feature transformation, in terms of both accuracy and completion time. The improvements over other techniques are substantial, particularly in the case of the accuracy of the clustering task. However, visualising data with very high dimensionality (i.e., greater than 100 dimensions) remains a challenge.

Person re-identification with block sparse recovery

We consider the problem of automatically re-identifying a person of interest seen in a “probe” camera view among several candidate people in a “gallery” camera view. This problem, called person re-identification, is of fundamental importance in several video analytics applications. While extracting knowledge from high-dimensional visual representations based on the notions of sparsity and regularization has been successful for several computer vision problems, such techniques have not been fully exploited in the context of the re-identification problem. Here, we develop a principled algorithm for the re-identification problem in the general framework of learning sparse visual representations. Given a set of feature vectors for a person in one camera view (corresponding to multiple images as they are tracked), we show that a feature vector representing the same person in another view approximately lies in the linear span of this feature set. Furthermore, under certain conditions, the associated coefficient vector can be characterized as being block sparse. This key insight allows us to design an algorithm based on block sparse recovery that achieves state-of-the-art results in multi-shot person re-identification. We also revisit an older feature transformation technique, Fisher discriminant analysis, and show that, when combined with our proposed formulation, it outperforms many sophisticated methods. Additionally, we show that the proposed algorithm is flexible and can be used in conjunction with existing metric learning algorithms, resulting in improved ranking performance. We perform extensive experiments on several publicly available datasets to evaluate the proposed algorithm.

Multi-Stage Recognition of Speech Emotion Using Sequential Forward Feature Selection

AbstractThe intensive research of speech emotion recognition introduced a huge collection of speech emotion features. Large feature sets complicate the speech emotion recognition task. Among various feature selection and transformation techniques for one-stage classification, multiple classifier systems were proposed. The main idea of multiple classifiers is to arrange the emotion classification process in stages. Besides parallel and serial cases, the hierarchical arrangement of multi-stage classification is most widely used for speech emotion recognition. In this paper, we present a sequential-forward-feature-selection-based multi-stage classification scheme. The Sequential Forward Selection (SFS) and Sequential Floating Forward Selection (SFFS) techniques were employed for every stage of the multi-stage classification scheme. Experimental testing of the proposed scheme was performed using the German and Lithuanian emotional speech datasets. Sequential-feature-selection-based multi-stage classification outperformed the single-stage scheme by 12–42 % for different emotion sets. The multi-stage scheme has shown higher robustness to the growth of emotion set. The decrease in recognition rate with the increase in emotion set for multi-stage scheme was lower by 10–20 % in comparison with the single-stage case. Differences in SFS and SFFS employment for feature selection were negligible.

Effectiveness of dereverberation, feature transformation, discriminative training methods, and system combination approach for various reverberant environments

The recently released REverberant Voice Enhancement and Recognition Benchmark (REVERB) challenge includes a reverberant automatic speech recognition (ASR) task. This paper describes our proposed system based on multi-channel speech enhancement preprocessing and state-of-the-art ASR techniques. For preprocessing, we propose a single-channel dereverberation method with reverberation time estimation, which is combined with multichannel beamforming that enhances direct sound compared with the reflected sound. In addition, this paper also focuses on state-of-the-art ASR techniques such as discriminative training of acoustic models including the Gaussian mixture model, subspace Gaussian mixture model, and deep neural networks, as well as various feature transformation techniques. Although, for the REVERB challenge, it is necessary to handle various acoustic environments, a single ASR system tends to be overly tuned for a specific environment, which degrades the performance in the mismatch environments. To overcome this mismatch problem with a single ASR system, we use a system combination approach using multiple ASR systems with different features and different model types because a combination of various systems that have different error patterns is beneficial. In particular, we use our discriminative training technique for system combination that achieves better generalization by making systems complementary with the modified discriminative criteria. Experiments show the effectiveness of these approaches, reaching 6.76 and 18.60 % word error rates on the REVERB simulated and real test sets. These are 68.8 and 61.5 % relative improvements over the baseline.