METHOD AND APPARATUS FOR AUDIO CLIP CLASSIFICATION

摘要

A method of classifying an audio clip into one of a plurality of predefined classes is disclosed. The method separates an audio clip into a plurality of segments, and each of the segments into a plurality of frames (102). The method then extracts Mel Frequency Cepstral Coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC) as the audio features from each frame within one segment (103). Segment characteristics for a segment are determined by deriving one or more measures of said audio features from each frame within the segment. Two or more Support Vector Machine (SVM) classifiers at two or more stages in a segment level classification process are utilised to determine a decision function value (104). For each segment of the audio clip, a class label is determined based on said decision function value. The decision function values of all the segments in said audio clip are mapped to associated segment confidence levels by using a sigmoid function model. Finally, post-processing (105) is performed on the classes and the confidence levels of all the segments to produce a class and associated confidence level for said audio clip (106).