A Weighted Approach Of Missing Data Technique Based On Sigmoid Function In Cepstral Domain

Recently the mainstream speech recognizer, based on Hidden Markov Model (HMM) and the Mel-frequency cepstral coefficients, has achieved good results in quiet environment. However, the performance deteriorates significantly in the presence of noise and other disturbance. As a result, many algorithms based on feature compensation or model adaptation, such as vector Taylor series, maximum linear likelihood regression, have been proposed to solve the problem.Missing data technique (MDT) has been proved a effective method in robust speech recognition. Conventionally, the identification of reliability is performed in spectral or log-spectral domain, and then the recognition is accomplished in log-spectral domain or cepstral domain followed by discrete cosine transformation. In order to apply MDT in cepstral domain, this paper presents a weighted average transformation to calculate the reliability of cepstral feature. The reliability is used as a weighting factor in the computation of Mahalanobis distance in HMM. Based on the additive noise model, the paper offers an origin about the weighted average transformation. It is deduced that the reliability enlarges the variance of HMM to reduce the mismatch between clean-training model and corrupted speech. This point is confirmed by the comparison of variances between in the clean-training model and in the multi-condition training model.To evaluate the performance of the propose algorithm, experiments are carried out on the Aurora2 speech database in the platform of HTK. The influence of parameters in the weighted MDT on the accuracy of recognition is analyzed carefully. and the algorithm is tested under the ideal condition and other two real condition. The result shows a distinct accuracy improvement compared to the baseline. Penalty factor and cooperation with other speech recognition methods are also tried in this paper. The main advantages of the weighted MDT are good interpretability, simple system implementation, low computation cost and easy to plug into other robust recognition algorithm. Finally, the research work is summarized.