Compensating the Speech Features via Discrete Cosine Transform for Robust Speech Recognition (基於離散餘弦轉換之語音特徵的強健性補償法)

基於離散餘弦轉換之語音特徵的強健性補償法 Compensating the speech features via discrete cosine transform for robust speech recognition Hsin-Ju Hsieh 謝欣汝, Wen-hsiang Tu 杜文祥, Jeih-weih Hung 洪志偉 暨南國際大學電機工程學系 Department of Electrical Engineering, National Chi Nan University Taiwan, Republic of China E-mail: s98323550@ncnu.edu.tw, aero3016@ms45.hinet.net, jwhung@ncnu.edu.tw Abstract In this paper, we develop a series of algorithms to improve the noise robustness of speech features based on discrete cosine transform (DCT). The DCT-based modulation spectra of clean speech feature streams in the training set are employed to generate two sequences representing the reference magnitudes and magnitude weights, respectively. The two sequences are then used to update the magnitude spectrum of each feature stream in the training and testing sets. The resulting new feature streams have shown robustness against the noise distortion. The experiments conducted on the Aurora-2 digit string database reveal that the proposed DCT-based approaches can provide relative error reduction rates of over 25% as compared with the baseline system using MVN-processed MFCC features. Experimental results also show that these new algorithms are well additive to many noise robustness methods to produce even higher recognition accuracy rates. I. Introduction Most of the state-of-the-art automatic speech recognition (ASR) system developed in the laboratory, in which the speech is not obviously distorted, can achieve excellent recognition per- formance. But in the real-world application, the recognition accuracy is seriously degraded due to so many distortions or variations existing in the application environment. Particularly speak- ing, the environmental distortions can be roughly classified into two types: channel distortion and additive noise, both influencing the performance of an ASR system a lot. The channel dis- tortion occurs when the speech signal is transmitted by electronic devices or transmission lines, such as the air, the telephone line or the microphone. The additive noise is like the “shadow” or “background” existing in the environment, such as car noise and babble noise. Noise robustness techniques have thus received much attention in recent years since they are so important in the applicability of ASR. One school of noise-robustness techniques is devoted to compensate the original speech fea- 211 ture to reduce the effect of noise and recover the speech feature back to its intact state. Typical examples of these techniques include cepstral mean normalization (CMN) [1], mean and vari- ance normalization (MVN) [2], cepstral gain normalization (CGN) [3], cepstral shape normaliza- tion (CSN) [4], histogram equalization (HEQ) [5], higher-order cepstral moment normalization (HOCMN) [6], temporal structure normalization (TSN) [7] and MVN plus ARMA filtering (MVA) [8]. However, the main purpose of the above methods can be roughly divided into two parts: one is to normalize the statistics of temporal-domain feature sequence and the other is to further reduce the mismatch by enhancing some components which are not easily affected by noise. For the latter case, the discrete Fourier transform (DFT) is usually used to be an analysis tool for obtaining the modulation spectrum of temporal-domain feature sequence. Therefore, we can deal with the modulation spectrum explicitly or implicitly in order to obtain the robust temporal-domain feature …

Content truncated to 500 words