Noise-Robust Speech Features Based on Cepstral Time Coefficients

Noise-Robust Speech Features Based on Cepstral Time Coefficients Ja-Zang Yeh Department of Science and Engineering National Sun Yat-Sen University Kaohsiung, Taiwan ycc97m@cse.nsysu.edu.tw Chia-Ping Chen Department of Science and Engineering National Sun Yat-Sen University Kaohsiung, Taiwan cpchen@cse.nsysu.edu.tw Abstract In this paper, we investigate the noise-robustness of features based on the cepstral time coef- ficients (CTC). By cepstral time coefficients, we mean the coefficients obtained from applying the discrete cosine transform to the commonly used mel-frequency cepstral coefficients (MFCC). Fur- thermore, we apply temporal filters used for computing delta and acceleration dynamic features to the CTC, resulting in delta and acceleration features in the frequency domain. We experiment with five different variations of such CTC-based features. The evaluation is done on the Aurora 3 noisy digit recognition tasks with four different languages. The results show all but one such feature set perfor- mance gain, the other feature sets actually lead to performance gains. The best feature set achieves an improvement of 25% over the baseline feature set of MFCC. Keywords: MFCC, CTC, delta, robust feature 1. Introduction A front-end of a speech recognition system may consist of several stages for noise-robustness to achieve good performance. In the early stage of spectral domain, well-known methods such as spectral subtraction [1] and Wiener filter [2] may be applied. In the middle stage of cepstral domain, the mel-frequency cepstral coefficients (MFCC) are commonly used as the static feature set. In the post- processing stage, there may be normalization, temporal information integration, and transformation modules. It has been observed that simple normalization approaches, such as the cepstral mean subtraction (CMS) [3], cepstral variance normalization (CVN) [4], and histogram normalization (HEQ) [5] can lead to significant performance improvement in recognition accuracy in noisy environment. Appar- ently such methods are capable of alleviating the mismatch between the clean and noisy data. In this paper we investigate novel features based on simple transformation methods. Specifically, we insert a window of static cepstral vectors in a matrix and then apply the discrete cosine transform (DCT) along the temporal axis. The coefficents after the DCT is called the cepstral time coefficients, 31 Figure 1: The block diagram of the proposed feature transformation methods. and the resultant matrix is called the cepstral time matrix (CTM) [6,7]. After CTM for each frame is extracted, we further apply normalization and routines for delta and acceleration feature extraction to the cepstral time coefficients. The transformed features are combined with the static MFCC features to form the final feature vector. This paper is organized as follows. Section 2 defines the cepstral time matrix and introduces the investigated feature transformations. The experimental setup and recognition results are described in Section 3. In Section 4, we draw conclusions. 2. Feature Transformations Our feature extraction and transformation process is illustrated in Figure 1. We begin with a review of the cepstral time matrix, which is followed by the mathematical definition of the proposed additive transformation methods. 2.1. Cepstral Time Coefficients We first insert a fixed number of adjacent feature …

Content truncated to 500 words