Alireza A. Dibazar

Feature analysis for automatic detection of pathological speech

This study focuses on a robust, rapid and accurate system for automatic detection of normal and pathological speech. This system employs noninvasive, non-expensive and fully automated measures of vocal tract characteristics and excitation information. Mel-frequency filterbank cepstral coefficients and measures of pitch dynamics were modeled by Gaussian mixtures in a hidden Markov model (HMM) classifier. The method was evaluated using the sustained phoneme /a/ data obtained from over 700 subjects of normal and different pathological cases from the Massachusetts Eye and Ear Infirmary (MEEI) database. This method attained 99.44% correct classification rates for discrimination of normal and pathological speech for sustained /a/. This represents 8% detection error rate improvement over the best performing classifier using carefully measured features prevalent in the state-of-the-art in pathological speech analysis.

Pathological Voice Assessment

While there are number of guidelines and methods used in practice, there is no standard universally agreed upon system for assessment of pathological voices. Pathological voices are primarily labeled based on the perceptual judgments of specialists, a process that may result in different label(s) being assigned to a given voice sample. This paper focuses on the recognition of five specific pathologies. The main goal is to compare two different classification methods. The first method considers single label classification by assigning a new label (single label) to the ensembles to which they most likely belong. The second method employs all labels originally assigned to the voice samples. Our results show that the pathological voice assessment performance in the second method is improved with respect to the first method

Cadence analysis of temporal gait patterns for seismic discrimination between human and quadruped footsteps

This paper reports on a method of cadence analysis for the discrimination between human and quadruped using a cheap seismic sensor. Previous works in the domain of seismic detection of human vs. quadruped have relied on the fundamental gait frequency. Slow movement of quadrupeds can generate the same fundamental gait frequency as human footsteps therefore causing the recognizer to be confused when quadruped are ambling around the sensor. Here we propose utilizing the cadence analysis of temporal gait pattern which provides information on temporal distribution of the gait beats. We also propose a robust method of extracting temporal gait patterns. Features extracted from gait patterns are modeled with optimum number of Gaussian Mixture Models (GMMs). The performance of the system during the test for discriminating between horse, dog, multiple people walk, and single human walk/run was over 95%.

The Application of Dynamic Synapse Neural Networks on Footstep and Vehicle Recognition

In this paper we report application of biologically based dynamic synapse neural network (DSNN) on perimeter protection. More specifically, the purpose is to protect a fence line from approaching human being and vehicles. We have used geophones to detect seismic signals generated by footsteps and vehicles. While acoustic sensors can be fooled by artificial sounds, fooling geophones by artificial seismic waves is a complicated task. Moreover detecting human footsteps -weak signal to noise ratio -by acoustic waveform is subject to the distance between the sensor and human. Therefore detecting a humans footsteps by employing acoustic information will not be possible unless he/she walks close to the acoustic sensors. Geophones are resonant devices; therefore any vibration in the substrate can generate seismic waveforms which could be very similar to the signature generated by footstep or vehicle. In addition, geophone response is completely substrate dependent, rendering recognition of footsteps or vehicle vs. other vibrations to be a very difficult task. Therefore, in order to have robust and high-confidence classification/detection of a human/ vehicle threats, we have employed the DSNN. The network is trained to extract intrinsic characteristics of the waveform, frame by frame. Then parameters of the network are analyzed by Gaussian mixture models. The results of our study show 88.8% and 86% correct classification rate for the detection of human footsteps and vehicle respectively.

Intelligent fence intrusion detection system: detection of intentional fence breaching and recognition of fence climbing

Perimeter fencing is widely used to isolate and protect public and private places such as airports, military bases, power stations and construction zones against unauthorized accesses. Fence structures merely prevent a percentage of intrusions or postpone them. A higher level of security is needed to monitor and investigate activities on or around the fences. The Fence Intrusion Detection System (FIDS) is one of Perimeter Intrusion Detection Systems (PIDS) focusing on the fence intrusions. In general, there is no system practically available to classify suspicious activities; whether the activity was due to the strong wind turbulent or climbing of a person on the fence.

Nonlinear Hebbian Learning for noise-independent vehicle sound recognition

In this paper we propose using a new approach, a nonlinear Hebbian learning, to implement acoustic signature recognition of running vehicles. The proposed learning rule processes both time and frequency components of input data. The spectral analysis is realized using auditory gammatone filterbanks. The gammatone-filtered feature vectors are then assembled over multiple temporal frames to establish a high-dimensional spectro-temporal representation (STR). With the exact acoustic signature of running vehicles being unknown, a nonlinear Hebbian learning (NHL) rule is employed to extract representative independent features from the spectro-temporal ones and to reduce the dimensionality of the feature space. During learning, synaptic weights between input and output neurons are adaptively learned. Motivated by neurobiological synaptic transmission in the brain, one specific nonlinear activation function, which can represent multiple independent neural signaling pathways, is proposed to process nonlinear Hebbian learning. It is shown that this function satisfies the requirements of the activation function in nonlinear neural learning, and that its derivative matches the implicit distribution of vehicle sounds, thus leading to a statistically optimal learning. Simulation results show that both STR and NHL can accurately extract critical features from original input data. The proposed model achieves better performance under noisy environments than its counterparts. For additive white Gaussian noise and common colored noise, the proposed model demonstrates excellent robustness. It can decrease the error rate to 3% with improvement 21 ~ 34% at signal-to-noise ratio (SNR)= 0 dB, and can function efficiently with error rate 7 ~ 8% at low SNR=-6 dB when its counterparts cannot work properly at this situation. To summarize, this study not only provides an efficient way to capture important features from high-dimensional input signals but also offers robustness against severe background noise.

Speaker recognition using dynamic synapse based neural networks with wavelet processing

Two problems in the field of speaker recognition are noise robustness and low interspeaker variability. This project involved the design of a system that is capable of speaker verification on a closed set of speakers using a wavelet processing technique that allows for a speaker‐dependent feature set extraction. Verification is accomplished using a dynamic synapse‐based neural network with noise‐resistance properties that is trained using a genetic algorithm technique. Using these techniques, the system was able to perform speaker verification without being adversely affected by normal levels of noise, and perform verification despite low variability between speakers.

Discrete Synapse Recurrent Neural Network for nonlinear system modeling and its application on seismic signal classification

For a lumped nonlinear modeling of the relationship between input and output sequences, Discrete Synapse Recurrent Neural Network (DSRNN) is proposed using fully Recurrent Neural Network (RNN) structure and Extended Kalman Filter (EKF) algorithm for its training. The training process is more efficient and there is less output error and more stability than in the previous study using feedforward networks. DSRNN is applied to a task of seismic signal classification to discriminate footsteps and vehicles from background. Temporal features of the signals were modeled using data recorded in the deserts of Joshua Tree, CA. The proposed classifier showed 0.3% false recognition rate for the recognition of human footsteps, 0.9% for vehicle, and 0.0% for background. The models were able to reject quadrupedal animals footsteps (in this study a trained dog). The system rejected dogs footsteps with 0.2% false recognition rate.

Protecting Military Perimeters from Approaching Human and Vehicle Using Biologically Realistic Dynamic Synapse Neural Network

The goal of this study is to detect and classify approaching human threats or vehicles, e.g. suicide bombers nearing a secured zone such as military bases. More specifically, this research is focused on (i) developing a vibration recognition system that can detect systematic vibration events; the entity might be a medium, human, animal, or a passenger vehicle, and (ii) discriminating between such a series of events vs. background and a single vibration event, e.g., falling of a tree limb. We have employed a seismic sensor to detect vibrations generated by footsteps and vehicles. A geophone is an inexpensive sensor which provides easy and instant deployment as well as long range detection capability. We have also designed a low power, low noise, and low cost hardware solution to process seismic waves locally where the sensor is located and wireless capability of the system makes it to communicate with a remote command center. Temporal features of the vibration signals were modeled by the dynamic synapse neural network (DSNN) using data recorded in the deserts of Joshua Tree, CA. The system showed 1.7% false recognition rate for the recognition of human footsteps, 6.7% for vehicle, and 0.0% for background. The models were able to reject quadrupedal animals footsteps (in this study a trained dog). The system rejected dogs footsteps with 0.02% false recognition rate.

The Gauss-Newton learning method for a generalized dynamic synapse neural network

A new architecture for dynamic synapse neural networks (DSNNs) has been introduced based on incorporating a continuous nonlinear mechanism to simulate synaptic neuro-transmitter release, adding a nonlinear output layer, and utilizing a Gauss-Newton learning method to train the network. We applied this network to simulate two nonlinear dynamical systems and then identify the dynamical systems by generating random noise observation data. The network estimation error per sample on the training phase was less than approximately 2% and on the test set was less than approximately 3%.