GMM With Modified Weight Applied Into Audio Segment Classification

In recent years the speech signal processing has become the frontier of signal processing field, Whether in speech recognition or retrieval, the audio is always pure speech,music or noise.While on the internet, the more data is mixed with different kinds of sounds together, in order to classify these data before speech recognition or retrieval, the audio classification technology was proposed.Based on the research of many related literatures in this field, the key of the technology can be summarized to two:one is to choose the features which statistically both have higher performance in distinguish and low computation for real-time speech processing requirements; Second, in the building of model, proposing new algorithm or changing original algorithm to improve the accuracy and efficiency of classification which make the system to achieve high stability.From this approach, in the selecting of features, considering time domain and frequency domain this paper achieved extraction of features in speech, music and noise documents separately.Based on the analysis of these characters' distribution in the frame levels and clip levels, finally choosing multidimensional feature-the ratio of sub-band energy which can accomplish classification requirements. In the model selection, because of the HMM model and the GMM model which derived from the HMM both can absorb the pronunciation of statistical acoustic properties in time of change. Therefore, in all kinds of existing classification systems, they have become the best recognition model. Also, because of GMM model unlike HMM that needs the divert probability of acoustic features in timing constraints, making GMM less calculation than HMM.From this point,GMM is more suitable for real-time processing. Based on the above consideration, this article selects GMM model for classification.Based on the further research in traditional GMM model. The paper proposed the GMM model with modified weight. Since the traditional GMM is training only by intra-class data, during classification, there may be some components in different GMM overlapping each other. That means the likelihoods of the overlap components are similar, which will induce the confusion during classification. To avoid this, the weight in GMM is modified according to the distance of components. For that component easily to confuse, the weight is deduced to decrease the effect of the component in whole result. Instead, for that component with well separability, the weight is enhanced.Because of the ratio of sub-band energy is only the best feature in frame level for GMM modeling,not clips levels.Besides this, the likelihood for judgement is based on each clip, which made the experiments of two kinds of models similar and difficult to distinguish. In order to make up this deficiency, an optimized likelihood is further put forward,which achieved the smooth processing of judgement in clips.From the experiments of F-measure based on the three systems, it can be drawn that GMM with modified weight has better performance than traditional GMM, and combined with optimized likelihood, the performance can be further enhanced.