Speech Recognition And Emotion Classification Algorithm Based On Spectral Context Feature

After years of development,speech recognition systems have had many promising applications,but their popularity is not ideal.In practical applications,speech recognition systems in specific environments are often difficult to achieve effective recognition.It is difficult to achieve satisfactory results from the interference with other factors which the speech recognition system was trained by laboratory environment,such as background noise,speaker's emotional changes,channel distortion,etc.Based on the key steps of speech recognition and emotion classification,and in light of the above research status and difficulties,this paper focuses on the research of recognition system in noisy environment and emotional environment.The main contents include:(1)Spectrum sequence context(SSC)feature is proposed.In the noise environment and different emotional environments,the traditional speech features are difficult to provide effective dynamic information.This paper presents a new feature extraction algorithm called SSC that is improved performance significantly,comparing with the traditional dynamic information processing methods.In the noise environment and speech emotion recognition have achieved good results.In other spectrum processes can be introduced to this feature,with a certain degree of universality.(2)To research the speech recognition in noisy environment,a speech recognition algorithm based on recursive plot compression distance is proposed.The speech signal is a timevarying sequence.By studying the significant function of the recursive plot model in the time series,the speech feature sequence is recursively compressed and the CK-1 distance between different recursive plots is calculated using the MPEG-1 compression algorithm.Complete the entire speech recognition process.(3)To research the classification decision in emotion recognition and propose a speech emotion recognition algorithm based on multi-feature deep confidence network.For different characteristics of different emotional description capabilities,we extract a variety of robust low-level features,and then feed it into the DBN network to extract high-level feature descriptors,and then use the idea of ensemble learning,ELM classification results vote fusion,in three kind of public dataset proves the effectiveness of this algorithm.