Research And Application Of Association Rules Mining In Speech Emotion Recognition

Using voice to deliver information is the most common and important way toexchange information for human beings’. Speech contains rich semantic andemotional information. People often only pay attention to the semantic informationcontained in speech, instead of the emotional information. Nevertheless, theseemotional information are critical for computer to understand human beings’emotions. In addition，as a significant research branch of affective computing, speechemotion is also an important research direction in the speech recognition field. Itrefers to psychology, artificial intelligence, signal processing, pattern recognition andother fields. Therefore, it is of great concern for research.Because speech emotion recognition has the dependency on voice signal features,different emotions associated with acoustic features. In this paper, we introduced theassociation rules algorithm to affective computing field for speech signal featureextraction. Then based on apriori, we proposed a prosody feature extraction algorithmin speech emotion (PFEA_AP algorithm) and conduct the corresponding experimenton Chinese emotion corpus and foreign emotion data set. The experimental resultsshowed that the features extracted by PFEA_AP algorithm not only reduced thedimension of emotion but also improved the classification accuracy. Therefore itverified the efficiency of our new algorithm and provided the theoretical basis offeature selection for future speech emotion recognition. The main work of this paperis as follows:First, analyze prosody feature and spectrum feature in speech emotionrecognition in details, for the complexity of speech emotion features, we used thesoftware praat to extract the feature data and discretize them, further to buildassociation rules mining mode for speech emotional prosodic features.Second, we extracted speech emotion prosodic features by using Apriorialgorithm and FP-growth algorithm. Then analyze and compare the advantages anddisadvantages of tow algorithm seperately from effective regular numbers producedby the algorithms and the running time.Eventually, we chose the algorithm of speechemotion rhythm feature extraction based on Apriori algorithm.Third, we conducted the simulation experiment by using SVM and BP neuralnetwork algorithm to extract features in terms of Chinese emotional corpus. Then wecompare the features selected separately using the fisher linear discriminant methodand PFEA_AP algorithm. It showed that our new algorithm is more effective. In orderto verify the emotional expression differences between people in different countries,we experimented in foreign data set EMO-DB. The results showed that differentcountries people emotional expression has not same pronunciation rules.