Thiago de M. Prego

A blind algorithm for reverberation-time estimation using subband decomposition of speech signals

An algorithm for blind estimation of reverberation time (RT) in speech signals is proposed. Analysis is restricted to the free-decaying regions of the signal, where the reverberation effect dominates, yielding a more accurate RT estimate at a reduced computational cost. A spectral decomposition is performed on the reverberant signal and partial RT estimates are determined in all signal subbands, providing more data to the statistical-analysis stage of the algorithm, which yields the final RT estimate. Algorithm performance is assessed using two distinct speech databases, achieving 91% and 97% correlation with the RTs measured by a standard nonblind method, indicating that the proposed method blindly estimates the RT in a reliable and consistent manner.

On the quality-assessment of reverberated speech

This paper addresses the problem of quantifying the reverberation effect in speech signals. The perception of reverberation is assessed based on a new measure combining the characteristics of reverberation time, room spectral variance, and direct-to-reverberant energy ratio, which are estimated from the associated room impulse response (RIR). The practical aspects behind a robust RIR estimation are underlined, allowing an effective feature extraction for reverberation evaluation. The resulting objective metric achieves a correlation factor of about 90% with the subjective scores of two distinct speech databases, illustrating the systems ability to assess the reverberation effect in a reliable manner.

On fault classification in rotating machines using fourier domain features and neural networks

The paper addresses the problem of classifying mechanical faults in rotating machines. In this context, three operational classes are considered, namely: normal (where the machine has no fault), unbalance (where the machine load has its weight not equally distributed), and misalignment (where the rotor and machine axes are dislocated from its natural concentric position). A large dataset consisting of 606 distinct scenarios is developed for system training and testing, along with a preprocessing strategy that improves data distribution among the three classes considered. A classifier based on an artificial neural network is described, achieving a global accuracy rate of 93.5%.

Feature analysis for quality assessment of reverberated speech

This paper analyzes the ability of several measurements to quantify the reverberation effect in speech signals. We consider an intrusive scheme, in which the clean and reverberated signals are available, allowing one to estimate the corresponding room impulse response (RIR) signal. An artificial neural network (ANN) is trained for all features and used in a regression approach to estimate the human perceptual evaluation in a mean opinion score (MOS) 1–5 scale. Dimensionality reduction approaches are applied to generate a simpler ANN regression, establishing the most representative features for the problem at hand. A correlation level of 85% with subjective test scores was achieved by reducing the input-vector dimension from 10 to 3, including only the features of reverberation time, room spectral variance, and direct-to-reverberant energy ratio.

The influence of feature vector on the classification of mechanical faults using neural networks

This paper investigates the problem of automatic detection of rotating-machine faults based on vibration signals acquired during machine operation. In particular, two new signal features, namely the kurtosis and entropy, are considered along with main spectral peaks to discriminate between several machine conditions: normal operation, (vertical and horizontal) misalignment, unbalanced load and bearing faults. Moreover, the inclusion of one set of three accelerometers for each roller bearing associated to the system acquiring more vibration signals also affects the generation of feature vector and is part of our proposal. In order to evaluate the rotating machine fault classification, a database of 1951 fault scenarios with several different fault intensities and rotating frequencies was designed and recorded, taking into consideration the specificities of the proposed machine learning task. The artificial neural networks recognition system employed in this work reached 95.8% of overall accuracy, showing the efficiency of the proposed approach.

Blind estimators for reverberation time and direct-to-reverberant energy ratio using subband speech decomposition

This paper describes algorithms for estimating two important features associated with the reverberation effect on speech signals: the reverberation time and direct-to-reverberant energy ratio. Both methods are referred to as blind algorithms in the sense that they are entirely based on the reverberant signal itself, not depending on the knowledge of the clean original signal. Proposed schemes use subband analysis to generate more and more reliable information, which is post-processed using basic statistical analysis to provide the desired estimate for each particular feature. Modifications on the original estimation algorithms are introduced to cope with lower SNRs. Performance of both algorithms is assessed under the ACE Challenge scope, which included a set of 288 speech signals for training and 4500 signals for final test. Results indicate the effectiveness of both techniques particularly in high-SNR situations.

Audio anomaly detection on rotating machinery using image signal processing

This paper addresses the problem of anomaly detection on rotating machinery in industrial environments using single channel audio signals. The proposed algorithm is based on image processing feature analysis obtained from the image representation of the Short-time Fourier Transform of reference and degraded audio signals. In order to assess the potential of the algorithm, a 8 signals database is recorded. The proposed algorithm is able to separate signals of machinery normal behavior from signals of machinery anomalous behavior with 100% hit rate using the recorded database.

Application of Machine Learning to Evaluate Unbalance Severity in Rotating Machines

This paper proposes two modifications in a classification method for unbalancing fault severity analysis in rotating machines based on the unbalancing mass force. The unbalancing severity was categorized into three severity levels, namely High (H), Medium (M) and Low (L). The feature vectors used information from discrete-time Fourier transform (DFT), kurtosis and entropy from the vibration signals. Similarity based Model (SBM) and Kernel discriminant analysis (KDA) techniques were applied in order to evaluate the feature discrimination and reduce the input feature space. All these techniques were tested in a random forest classifier. Test results indicate that non-linear transformations to the feature space combined to random forest can further improve the classification of unbalancing severity defect, by reducing the feature space dimension from 31 to 6.

Application of Machine Learning in Diesel Engines Fault Identification

The objective of this work is the fault diagnosis in diesel engines to assist the predictive maintenance, through the analysis of the variation of the pressure curves inside the cylinders and the torsional vibration response of the crankshaft. Hence a fault simulation model based on a zero-dimensional thermodynamic model was developed. The adopted feature vectors were chosen from the thermodynamic model and obtained from processing signals as pressure and temperature inside the cylinder, as well as, torsional vibration of the engines flywheel. These vectors are used as input of the machine learning technique in order to discriminate among several machine conditions, such as normal, pressure reduction in the intake manifold, compression ratio and amount of fuel injected reduction into the cylinders. The machine learning techniques for classification adopted in this work were the multilayer perceptron (MLP) and random forest (RF).

On the enhancement of dereverberation algorithms using multiple perceptual-evaluation criteria

This paper describes an enhancement strategy based on several perceptual-assessment criteria for dereverberation algorithms. The complete procedure is applied to an algorithm for reverberant speech enhancement based on single-channel blind spectral subtraction. This enhancement was implemented by combining different quality measures, namely the so-called QAreverb, the speech-to-reverberation modulation energy ratio (SRMR) and the perceptual evaluation of speech quality (PESQ). Experimental results, using a 4211-signal speech database, indicate that the proposed modifications can improve the word error rate (WER) of speech recognition systems an average of 20%.