Khairul Anuar Ishak

Speaker Verification using Vector Quantization and Hidden Markov Model

This paper presents a speaker verification system using a combination of vector quantization (VQ) and hidden Markov model (HMM) to improve the HMM performance. A Malay spoken digit database which contains 100 speakers is used for the testing and validation modules. It is shown that, by using the proposed combination technique, a total success rate (TSR) of 99.97% is achieved and it is an improvement of 11.24% in performance compared to HMM. For speaker verification, true speaker rejection rate, impostor acceptance rate and equal error rate (EER) are also improved significantly compared to HMM.

A Speed limit Sign Recognition System Using Artificial Neural Network

This paper presents a real-time system to detect speed limit signs and remind drivers about the allowable speed limit in a specific road. The developed system consists of two main tasks, namely detection and recognition. In our work, speed limit sign is detected and extracted from real world scenes on the basis of their color and shape features. The detection task is based on a combination of color segmentation and shape detection techniques. It significantly speeds up the shape detection process by calculating the cross-correlation in frequency domain. Next, classification is then performed on extracted candidate region using multi-layer perceptron neural networks. Experiment results proved the feasibility of this system.

Decision Fusion for Isolated Malay Digit Recognition Using Dynamic Time Warping (DTW) and Hidden Markov Model (HMM)

This paper is focused on Malay speech recognition with the intention to introduce a decision fusion technique for isolated Malay digit recognition using Dynamic Time Warping (DTW) and Hidden Markov Model (HMM). This study proposes an algorithm for decision fusion of the recognition models. The endpoint detection, framing, normalization, Mel Frequency Cepstral Coefficient (MFCC) and vector quantization techniques are used to process speech samples to accomplish the recognition. Decision fusion technique is then used to combine the results of DTW and HMM. The algorithm is tested on speech samples that is a part of a Malay corpus.

Normalized Least Mean Square adaptive noise cancellation filtering for speaker verification in noisy environments

In this paper, we present a speaker verification system based on the hidden Markov models (HMMs) and normalized least mean square (NLMS) adaptive filtering. The aim of using NLMS adaptive filtering is to improve the HMMs performance in noisy environments. A Malay spoken digit database is used for the testing and validation modules. It is shown that, in a clean environment a total success rate (TSR) of 89.97% is achieved using HMMs. For speaker verification, the true speaker rejection rate is 25.3% while the impostor acceptance rate is 9.99% and the equal error rate (EER) is 16.66%. In noisy environments without NLMS adaptive filtering TSRs of between 43.07%-51.26% are achieved for SNRs of 0-30 dBs. Meanwhile, after NLMS filtering, TSRs of between 55.18%-55.30% are achieved for SNRs 0-30 dB.

Evaluation of face recognition system using Support Vector Machine

Face recognition is an interest subject in pattern recognition study for machine learning applications. It is a non-intrusive system which requires minimal participation from user in order to perform identification tasks. In this paper we present a face recognition system based on Support Vector Machine (SVM) which acts as a multiclass classifier. The performance of this system is evaluated using Yale database with various facial expressions and illumination conditions. This method train and test the images with raw image data of 625 features. The result has achieved an encouraging recognition rates compares to Principal Component Analysis method (PCA).

Robust digit recognition with dynamic time warping and recursive least squares

Robustness is a key issue in speech recognition. This paper proposes a speech recognition algorithm for Malay digits from 0 to 9. This paper also proposes an algorithm for noise cancellation by using recursive least squares (RLS). This system consists of speech processing inclusive of digit margin and recognition which uses zero crossing and energy calculations. Mel-Frequency Cepstral Coefficient (MFCC) vectors are used to provide an estimate of the vocal tract filter. Meanwhile dynamic time warping (DTW) is used to detect the nearest recorded voice with appropriate global constraint. The global constraint is used to set a valid search region because the variation of the speech rate of the speaker is considered to be limited in a reasonable range, which means that it can prune the unreasonable search space. The algorithm is tested on speech samples that are recorded as a part of a Malay corpus. The results show that the algorithm can recognize almost 80.5% of the Malay digits for all recorded words. By adding RLS noise canceller in the preprocessing stage it increases the accuracy to 92.3%.

Enhancing speaker verification in noisy environments using Recursive Least-Squares (RLS) adaptive filter

In this paper, we present a speaker verification system based on the Hidden Markov Model (HMM) technique and Recursive Least Squares (RLS) adaptive filtering. The aim of using RLS adaptive filtering is to improve the HMM performance in noisy environments. A Malay spoken digit database is used for the testing and validation modules. It is shown that, in a clean environment a total success rate (TSR) of 89.97% is achieved using HMM. For speaker verification, the true speaker rejection rate is 25.3% while the impostor acceptance rate is 9.99% and the equal error rate (EER) is 16.66%. In noisy environments without RLS adaptive filtering TSRs of between 43.07%–51.26% are achieved for SNRs of 0–30 dBs. Meanwhile, after RLS filtering, TSRs of between 50.95%–56.75% are achieved for SNRs 0–30 dB.

Improving hybrid speaker verification in noisy environments using least mean-square adaptive filters

In this paper, we present a hybrid speaker verification system based on the Hidden Markov Models (HMMs) and Vector Quantization(VQ) and Least Mean-Square (LMS) adaptive filtering. The aim of using hybrid speaker verification is to improve the HMMs performance, while LMS adaptive filtering is to improve the hybrid speaker verification performance in noisy environments. A Malay spoken digit database is used for the training and testing. It is shown that, in a clean environment a Total Success Rate (TSR) of 99.97% is achieved using hybrid VQ and HMMs. For speaker verification, the true speaker rejection rate is 0.06% while the impostor acceptance rate is 0.03% and the equal error rate (EER) is 11.72%. In noisy environments without LMS adaptive filtering TSRs of between 62.57%-76.80% are achieved for Signal to Noise Ratio (SNR) of 0-30 dBs. Meanwhile, after LMS filtering, TSRs of between 77.31%-76.87% are achieved for SNRs of 0-30 dB.

A non linear face recognition system using Support Vector Machine

A face recognition system uses face to verify individuals using computing capability. However, its performances often degrade due to high dimensional data and large feature appearance of the face image. This paper present a face recognition system based on non linear feature extraction technique to reduce the dimensionality of the face image, called Locally Linear Embedding. This method considers the hidden layer of face manifold to be the input of a SVM multiclass classifier. The performance is evaluated using the ORL database and achieved better recognition rates than the Principal Component Analysis.

Investigation on Different Pre-processing Approaches for Face Recognition System

One of the challenges in face recognition system is to deal with inhomogeneous intensity problem that occur with different lighting conditions. In this paper, comparisons are made on several pre-processing methods i.e. histogram equalization, local binary pattern, wavelet transform and multiscale retinex. First, the input image is pre-processed with the illumination correction method before the classification task is done. The results are evaluated using the Yale, ORL and our own UKM database. These databases include images with various illumination conditions and expressions. Using PCA as the feature extraction and Euclidean Distance as the classification purposed, our experiments shows that multiscale retinex achieved the lowest equal error rates with 5.03% followed by local binary pattern (7.52%), wavelet transform (12.53%) and histogram equalization (12.97%) on average for all three databases.