Wavelet Energy-Based Support Vector Machine for Noisy Word Boundary Detection With Speech Recognition Application

Wavelet Energy-Based Support Vector Machine for Noisy Word Boundary Detection With Speech Recognition Application Chia-Feng Juang, Chun-Nan Cheng and Chiu-Chuan Tu Department of Electrical Engineering National Chung-Hsing University, Taichung, 402 Taiwan, R.O.C. e-mail: cfjuang@dragon.nchu.edu.tw Abstract Word boundary detection in variable noise-level environments by support vector machine (SVM) using Low-band Wavelet Energy (LWE) and Zero Crossing Rate (ZCR) features is proposed in this paper. The Wavelet Energy is derived based on Wavelet transformation; it can reduce the affection of noise in a speech signal. With the inclusion of ZCR, we can robustly and effectively detect word boundary from noise with only two features. For detector design, a Gaussian-kernel SVM is used. The proposed detection method is applied to detection word boundaries for an isolated word recognition system in variable noisy environments. Experiments with different types of noises and various signal-to-noise ratios are performed. The results show that using the LWE and ZCR parameters-based SVM, good performance is achieved. Comparison with another robust detection method has also verified the performance of the proposed method. Keywords: Speech detection, word boundary detection, support vector machine, wavelet transform, noisy speech recognition. 1. INTRODUCTION For speech recognition, the detection of speech affects recognition performance. A robust word boundary detection method in the presence of variable-label noises is necessary and is studied in this paper. Depending on the characteristics of speech, a variety of parameters have been proposed for boundary detection. They include the time energy (the magnitude in time domain), zero crossing rate (ZCR) [1] and pitch information [2]. These parameters usually fail to detect word boundary when signal-to-noise ratio (SNR) is low. Another parameter concerning frequency domain has also been recently proposed. According to the frequency energy, the time-frequency (TF) parameter [3] which sums the energy in time domain and the frequency energy was presented. The TF-based algorithm may work well for fixed-level background noise. However, its detection performance degrades for background noise of various levels. For this problem, some modified TF parameters are proposed [4]. In [5], the idea of using Wavelet transform features as speech detection features was proposed. In this paper, we present a new Low-band Wavelet Energy (LWE) parameter which separates the speech from noise in the domain of Wavelet transform. Computation of the WE parameter is easier than the modified TF parameters, and it is shown in the experiment section that a better detection performance is achieved. 21 After the features for detection have been extracted, the next step is to determine thresholds and decision rules. Many decision methods based on computational intelligence techniques have been proposed, such as fuzzy neural networks (FNNs) [4] and neural networks (NNs) [6]. Generalization performance may be poor when FNNs and NNs are over-trained. To cope with the low generalization ability problem, a new learning method, the Support Vector Machine (SVM), has been proposed [7, 8]. SVM is a new and useful learning method whose formulation is based on the principle of structural risk minimization. Instead of minimizing an objective function based on training, …

Content truncated to 500 words