Robust Hierarchical Linear Model Comparison forUtterance-End Detection Under Noisy Environments

• Main Authors: Ming-Chaing Hsu, 徐明江
• Other Authors: Wei-Tyng Hong
• Format: Others
• Language: en_US
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/55990173986262954692

碩士 === 元智大學 === 通訊工程學系 === 99 === To detect speech-block end-points, we use entropy of the speech signal, and compare the t of two weighted linear models for the entropy. The regression models are constructed so that their t will di er the most near the speech-block end-points. For estimation of speech signal entropy we use a histogram of speech signal that is sampled in a frame of xed duration. The resulting sequence of entropies in consecutive frames is used for tting the linear models. The models are tted in a sliding interval of the entropies, that correspond to several consecutive frames in a sliding time window. The interval with the greatest di erence of the model t is used to estimate the location of the speech-block boundaries. Model M1 is very simple; it corresponds to a constant average entropy level for the speech signal in the entire window. Model M2 models step-like entropy change from one constant level to another, with a gradual transition between the levels. It is a piecewise linear regression model with two horizontal lines connected by a third transitional line. We treat M1 as a linear model only to be able to describe it as a sub-model of M2 and use statistical methodology for sub-model testing. The performance of the presented algorithm is compared with a standard EOU (end-of-utterance) detection. The average rate of proper detection is 92.36% for our approach. It gets almost 13% improvement over the EOU method. Several tables are presented to illustrate that, while obtaining similar result for clean sound, our regression model approach has better ability to resist the negative e ect of noise.