Showing papers by "Yaniv Zigel published in 2010"

Nocturnal sound analysis for the diagnosis of obstructive sleep apnea

Abstract: A novel method for screening obstructive sleep apnea syndrome (OSAs) based on nocturnal acoustic signal is proposed. Full-night audio signals from sixty subjects were segmented into snore, noise and silence events using semi-automatic algorithm based on Gaussian mixture models which achieves more than 90% (92%) sensitivity (specificity) and produces an average of 2,000 snores per subject. A classification into 3 groups is proposed for the diagnosis: comparison group - non-OSA subjects (apnea hypopnea index, AHI 30). A Bayes classifier was implemented, fed with five acoustic features, all correlated with the severity of the syndrome: (1) Inter Event Silence, which quantifies segments suspicious as apnea; (2) Mel Cepstability, measures the entire night stability of the spectrum, expressed using mel-frequency cepstrum; (3) Energy Running Variance, a criterion for the variation of the nocturnal acoustic pattern; (4) Apneic Phase Ratio, exploiting the finding that snores around apnea events expressing larger acoustic variation; and (5) Pitch Density. Correct classification of 92% for resubstitution method and 80% for 5-fold cross validation method was achieved. Moreover, in a case of two groups with a threshold of AHI=10, a sensitivity (specificity) of 96.5% (90.6%) and 87.5% (82.1%) for resubstitution and cross-validation respectively were obtained.

Age recognition based on speech signals using weights supervector.

Abstract: This paper proposes a new age-recognition system approach— building a Gaussian mixture model–based weights supervector features for a support vector machine (SVM). This approach uses the hypothesis that it is possible to find unique Gaussians for each age-group model in the universal background model (UBM). The weights of those Gaussians can lead to a discriminant way to separate the age groups. The suggested approach was tested on two corpora (aGender and local corpus) with classification into four age groups, achieving 53.75% and 56.18% weighted average recall, respectively, which are better results compared to the state-of-the-art classifier.

Room volume classification from room impulse response using statistical pattern recognition and feature selection

Abstract: Classification of the room volume from the room impulse response (RIR) can be useful in acoustic scene analysis applications, using RIR that is provided directly, or estimated from audio recordings. Current methods for estimating the room volume from the RIR require the source-to-receiver distance, and may be sensitive to differences in absorption. A room volume classification method is presented that does not require the source-to-receiver distance, and which is potentially robust to differences in absorption. Room volume features are defined that are related to the room volume and may be extracted from the RIR. Gaussian mixture models are trained to model room volume classes. Room volume is classified according to a maximum likelihood criterion that is normalized with a background model. Feature selection is performed with different classification error criteria. Both simulated and measured RIRs were examined, achieving an equal error rate of 0.1% and 19.1%, respectively.

Method and system for detecting p-waves in the surface ecg signal

Abstract: The invention provides an ECG system and apparatus for detecting P-waves even in the patients with arrhythmia. The system is capable of obtaining sufficient data from eight leads and to display an ECG construct with marked or emphasized P-waves.

The Effect of Reverberation on Optimal GMM Order and CMS Performance in Speaker Verification Systems

Abstract: In speaker recognition, features are extracted from speech signals to form feature vectors, and statistical pattern recognition methods are applied in order to model the distribution of the feature vectors in the feature space. Speakers are recognized by pattern matching of the statistical distribution of their feature vectors with target models. Speaker verification (SVR) is the task of deciding, upon receiving tested feature vectors, whether to accept or reject a speaker hypothesis, according to the speaker’s model. A popular feature extraction method for speech signal processing is the mel-frequency cepstral coefficients (MFCC) [Davis & Mermelstein, 1980], and Gaussian mixture models (GMM) has become a dominant approach for statistical modeling of speech feature vectors for text-independent SVR [Reynolds et al., 2000]. Speaker verification is widely used in telecommunication or conference room applications, where reverberation is often present due to the surrounding room environment. The presence of reverberation adds distortion to the feature vectors, which results in performance degradation of SVR systems due to mismatched conditions between trained models and test segments. Feature normalization techniques such as the cepstral mean subtraction (CMS) [Mammone et al., 1996] and variance normalization [Chen & Bilmes, 2007], and score normalization techniques such as the Znorm, Hnorm, Tnorm [Bimbot et al., 2004, Mammone et al., 1996] and Top-norm [Zigel & Wasserblat, 2006], were originally developed to compensate for the effect of a telephone channel [Mammone et al., 1996], or for the effect of slowly varying convolutive noises in general [Reynolds et al., 2000]. For that reason, these techniques may be used to reduce the effect of reverberation, if it is characterized by a short-duration room impulse response (RIR). However, it may be difficult to find research studies in the literature on the effect of CMS on SVR performance under reverberation conditions of long duration RIR, which is often the case in room acoustics. In cases of long-duration RIR, the target models may be trained using a reverberant speech database, as suggested by Peer et al. [Peer et al., 2008], in order to overcome the mismatched conditions between the models and the reverberant testing speech segments. This method was tested on adaptive-GMM (AGMM) based SVR system, with various values of reverberation time (RT the time that takes the impulse response to decay by 60dB [Schroeder,