Dominic Mathew

Phoneme classification using Reservoirs with MFCC and Rasta-PLP features

The performance of speech classification tasks can be improved by accurate acoustic modeling. This modelling is responsible for establishing the relationship between the speech signal and the phonetic units that were produced by the speaker. In this paper Acoustic Modeling(AM) is done using Reservoir Computing(RC) technique for which the input speech signal frames can be identified and classified among the 39 different phoneme classes. The speech samples are taken from Timit database. Mel-frequency cepstral coefficient(MFCC) and Relative spectral transformation Perceptual Linear Prediction(Rasta-PLP) coefficients are used as features. We compared the performance of the reservoir system using MFCC and Rasta-PLP coefficient extraction methods.

Speaker identification and verification of noisy speech using multitaper MFCC and Gaussian Mixture models

The two major applications of speaker recognition applications are speaker verification and speaker identification. But in most of the cases the signal is corrupted with background interferences such as noise and echo. This paper proposes the method of speaker recognition and identification after the noise separation. Support Vector Machine(SVM) classification based signal separation is adopted here. MFCC and Multitaper MFCC are used for feature extraction. Despite having low bias, MFCC has large variance. One promising technique for reducing the variance is to replace Hamming windowed spectrum with a multi-taper spectrum estimate. Gaussian Mixture models along with Universal Background Model(UBM) is used for speaker verification and identification tasks.

A novel approach for contrast enhancement and noise removal of medical images

This paper presents the application of two different image enhancement techniques to medical images and the comparison of these techniques with traditional Histogram Equalization (HE) method. The proposed method uses Weighted Histogram Equalization (WHE) and transform domain approach to enhance medical images. Simulation results shows that Perona-Malik filter (PM filter) can be used to remove the noise from the enhanced image without impacting the image contrast. Peak Signal to Noise Ratio (PSNR) values shows that the use of Perona-Malik filter can improve both the methods. The proposed WHE method with PM filter shows better PSNR values than the transform domain approach with PM filter.

Person verification using multimodal biometric system

This paper presents a multimodal approach for person verification based on the features extracted from signature, face and iris of an individual. Features from signatures are extracted using Discrete Cosine Transform (DCT) and by applying Sparse Representation techniques. Facial features are extracted using Gabor Filter bank and Kernel Principal Component Analysis (KPCA). In this work, for extracting features from iris, we proposed Gabor filter bank and KPCA. The feature vectors so obtained are then given as input to classifiers. Support Vector Machines (SVM) classifiers are used for the three modalities. The final decision of multimodal system is based on the majority voting of classifiers. The SVM classifiers are trained and tested using the following databases — SUSIG-Signature, ORL-Face and UBIRIS — Iris. The experimental analysis shows that the performance of multimodal system has attained a GAR of 99.5% at an FAR of 0%.

Steganographic Technique Using Secure Adaptive Pixel Pair Matching for Embedding Multiple Data Types in Images

Steganography is one of the newly emerging area of development in the field of signal processing. It is the technique of hiding the fact that communication is even taking place, by hiding information in any other information. In this paper, a secure data-hiding method is proposed for hiding multiple datatypes such as text, image and audio within images by Secure Adaptive Pixel Pair Matching(SAPPM) which can be effectively used for colour images as well as gray scale images. The secret message is encrypted using a pseudo random sequence, which is generated according to the secret key given by the user. The encrypted secret message is hidden in cover image using SAPPM algorithm. Compared with previous methods like Least Significant Bit(LSB) substituition, simple Diamond Encoding(DE) and APPM, the proposed method provides a much more stronger algorithm which incorporates cryptography along with steganography, and always have a lower distortion.

Acoustic modeling using auditory model features and Convolutional neural Network

The state of art automatic speech recognition systems use Deep Neural Networks(DNN) for acoustic modeling. More recently, Convolutional neural Networks(CNN) have shown substantial acoustic modelling capabilities due to its ability to deal with structural locality in the feature space. In this paper, a detailed study of CNN based acoustic models on TIMIT database has been performed. For feature extraction an biologically motivated auditory model is simulated using Patterson and Holdsworth filter bank. MFSC features are also extracted for comparison. The experiments show that CNN with the auditory model features outperforms the conventional acoustic models which use mel spectral features.

A Novel Acoustic Fingerprint Method for Audio Signal Pattern Detection

This paper presents a novel and efficient audio signal recognition algorithm with limited computational complexity. As the audio recognition system will be used in real world environment where background noises are high, conventional speech recognition techniques are not directly applicable, since they have a poor performance in these environments. So here, we introduce a new audio recognition algorithm which is optimized for mechanical sounds such as car horn, telephone ring etc. This is a hybrid time-frequency approach which makes use of acoustic fingerprint for the recognition of audio signal patterns. The limited computational complexity is achieved through efficient usage of both time domain and frequency domain in two different processing phases, detection and recognition respectively. And the transition between these two phases is carried out through a finite state machine(FSM)model. Simulation results shows that the algorithm effectively recognizes audio signals within a noisy environment.

Realization of energy harvesting system using a frequency generator

This paper presents the hardware implementation of a frequency converter and AC/DC step up converter for realising the low voltage energy harvesting. In the simulation part, the input of the AC/DC converter is from an electromagnetic micro generator which converts vibration into electrical energy. The electromagnetic micro generator is being modelled and simulated. But in the hardware implementation, a frequency generator has been designed to generate the output of micro generator. The AC/DC converter is controlled by ARM7 microcontroller. A closed loop control is incorporated.