Sub-banded reconstructed phase spaces for speech recognition

Indrebo, Kevin M.; Povinelli, Richard J.; Johnson, Michael T.

doi:10.1016/j.specom.2004.12.002

Cited by 17 publications

(7 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We have shown in previous work that even when these assumptions are not met RPSs contain important information for classifying a signal [9]. Here these signals are the 12 lead ECGs, and the classes are infarcted or non-infarcted.…”

Section: Rps/gmm Approachmentioning

confidence: 98%

See 1 more Smart Citation

RPS/GMM Approach toward the localization of myocardial infarction

Mneimneh

Povinelli

2007

2007 Computers in Cardiology

View full text Add to dashboard Cite

IntroductionDue to the disconnections between clinical scientists and applications of the available diagnosing methods, the current methods for detecting ischemic heart disease are not as well developed as therapeutic interventions despite the knowledge regarding the underlying pathophysiology. Thus, the 2007 Physionet/Computers in Cardiology Challenge embraces the disciplines of both electrocardiography and Magnetic Resonance Imaging (MRI). However in this work we propose the use of the 12 leads ECG signals to locate, according to the standard 17 segments [1], the infarcted segments, the centroid of the infarction, and the extent of the infarct. The results of the challenge are compared to gold standards taken from MRI with gadolinium enhanced images. A Reconstructed Phase Space (RPS)/ Gaussian Mixture Model (GMM) [2] signal modeling approach is used to identify the infarcted segments.The paper is organized in the following sections: background, data set, method, results, discussion and conclusion. BackgroundThe American Heart Association [3] shows how the 12 ordinary leads can be related back to infarction locations. The drawback in the provided rules is that they do not relate to the 17 segment locations provided in this challenge. In [1], de Luna and Wagner provided myocardial infarction localization rules that relate the Q wave, R wave, and S wave to the 17 wall segments using MRI.Several research groups have used ECG measurements to locate infarcts. These include the work by Reddy et al. In [4], they developed an approach based on QRS measurement and neural networks to classify healthy and patients with myocardial infarction. Their accuracy was 79% with a specificity of 97%.Lu et al's neural-fuzzy approach for classifying myocardial infarction uses ST elevation as an input parameter. Their approach identified whether a lead is infracted, and from the infarcted lead set inferred the infarct location. Their accuracy for detecting healthy patients was 89.4% and for detecting infarcted patients was 95.0% [5]. DatasetsThe challenge dataset consists of two training and two testing patient records. One record has a moderate to large infarct, while the second has a relatively compact infarct. This data is labor-intensive to produce. The Body Surface Potential Map data, consisting of ECG data for 352 torsosurface sites, is provided for an averaged PQRST complex signal sampled at 2kHz. The 12 ordinary leads and the Frank leads are also provided [6].However, this dataset is not sufficient for building a classification method. Therefore, we supplement the current challenge dataset with the PTB Diagnostic ECG Database, which provides a dataset of healthy and infarcted patients that is used as a training set for learner our classifier. The PTB database contains 549 records from 294 subjects. From the 549 records, 367 records taken from 148 patients had myocardial infarction [7]. MethodThe RPS/GMM approach has been applied previously in the detection of myocardial ischemia [2]. Here, we apply the RPS/GMM approac...

show abstract

Section: Rps/gmm Approachmentioning

confidence: 98%

“…GMMs are a set of Gaussian probability density functions used to characterize the distribution of an underlying set of data. They are widely used in engineering applications, especially speech [9]. The equation that defines a GMM is [2]:…”

Section: Rps/gmm Approachmentioning

confidence: 99%

RPS/GMM Approach toward the localization of myocardial infarction

Mneimneh

Povinelli

2007

2007 Computers in Cardiology

View full text Add to dashboard Cite

show abstract

“…In [17], reconstructed phase spaces built from speech signals that have been subband filtered were used for isolated phoneme classification, showing improved recognition accuracies over fullband signal phase space reconstruction features. However, this approach is infeasible for continuous speech recognition because of its high computational complexity.…”

Section: Nonlinear Features For Speech Recognitionmentioning

confidence: 99%

Third-Order Moments of Filtered Speech Signals for Robust Speech Recognition

Indrebo

Povinelli

Johnson

2006

Nonlinear Analyses and Algorithms for Speech Processing

Self Cite

View full text Add to dashboard Cite

Novel speech features calculated from third-order statistics of subband-filtered speech signals are introduced and studied for robust speech recognition. These features have the potential to capture nonlinear information not represented by cepstral coefficients. Also, because the features presented in this paper are based on the third-order moments, they may be more immune to Gaussian noise than cepstrals, as Gaussian distributions have zero third-order moments. Preliminary experiments on the AURORA2 database studying these features in combination with Mel-frequency cepstral coefficients (MFCC's) are presented, and improvement over the MFCC-only baseline is shown with the combined feature set for several noise conditions.

show abstract

“…Unfortunately, such a model cannot convey nonlinear 3D fluid dynamics phenomena of speech [20,21]. In order to fill the existing gap between this ideal linear deterministic model and real strongly unpredictable speech production process, non-linear processing techniques can be used [22,23,24,25].In recent years, reconstructed phase space (RPS) of speech has been used for speech recognition [26,27], speech enhancement [26,27]and detecting sleepiness [30]. Moreover, nonlinear dynamics features extracted from RPS of speech has been employed in SER.It has been shown that geometrical properties of RPScontain important emotional cues of speaker [24,25].…”

Section: Introductionmentioning

confidence: 99%

Recognition Of Emotion Using Reconstructed Phase Space Of Speech

Harimi

Fakhr

Bakhshi

2016

MJCS

View full text Add to dashboard Cite

In recent years, automatic recognition of human's emotion from speech has become one of the most important research areas, which can improve man-machine interaction. In this study, we proposed new features derived from reconstructed phase space (RPS) of speech. To this end, the RPS is uniformly divided into non-overlapping discrete cells and the number of points included in each cell is counted to form the proposed feature vector. Then multiple classifiers were examined to classify speech samples according to their emotional states. Our experimental results have demonstrated the potential and promise of proposed RPS based features as a useful combination for standard prosodic and spectral features. The best average recognition rate of 89.34% was obtained for classifying seven emotion categories in the Berlin database using a support vector machine with both radial basis function and polynomial kernels.Keywords: speech emotion recognition; reconstructed phase space. INTRODUCTIONSpeaking is the fastest and one of the most important communication means of human beings. Now a days, automatic speech recognition (ASR) systems are extensively used in man-machine interaction. However, human speech is generally embedded with emotions to convey the intended message. From this fact arises a new multidisciplinary research area known as Speech emotion recognition (SER). Despite of widespread efforts done in SER, there are many challenging problems that need to be solve in order to improve the performance of SER systems [1]. Based on contradictory reports on the effect of emotions on some of acoustic attributes, specifying effective features is still the major unsolved problem.To solve the above problem, many prosodic and spectral features have been presented. Prosodic features appear when sounds are put together in connected speech and they are mainly deal with intonation, stress and rhythm of speech. It has been shown that prosodic features, which are widely used in SER, suggest important emotional cues of the speaker [1][2][3][4][5][6][7]. In theliterature, pitch, duration, energy and their derivatives are widely used to representprosodic features [1,8]. Features extracted from the spectrum of speech are generally called spectral. These features convey the frequency contents of the signal and provide complementary information for prosodic features Both prosodic and spectral features are generally computed by the traditional linear source-filter model of the human speech production system [19]. Unfortunately, such a model cannot convey nonlinear 3D fluid dynamics phenomena of speech [20,21]. In order to fill the existing gap between this ideal linear deterministic model and real strongly unpredictable speech production process, non-linear processing techniques can be used [22,23,24,25].In recent years, reconstructed phase space (RPS) of speech has been used for speech recognition [26,27], speech enhancement [26,27]and detecting sleepiness [30]. Moreover, nonlinear dynamics features extracted from RPS of speech has b...

show abstract

Sub-banded reconstructed phase spaces for speech recognition

Cited by 17 publications

References 43 publications

RPS/GMM Approach toward the localization of myocardial infarction

RPS/GMM Approach toward the localization of myocardial infarction

Third-Order Moments of Filtered Speech Signals for Robust Speech Recognition

Recognition Of Emotion Using Reconstructed Phase Space Of Speech

Contact Info

Product

Resources

About