A Combination of Data Mining Method with Context-Based State Transfer for Speech/Music Discrimination

In our previous work, a speech/music classifier is proposed on the basis of the feature subset selection (FSS) tool and oblique decision tree induced by the algorithm OC1. In this paper, we endeavor to improve it by state transfer (ST) strategy whose aim is to refine the classification results, according to the fact that adjacent segments in one audio file have strong relevance to each other. The proposed algorithm is evaluated by a set of 5-to-11-minute 504 audio files of different types of speech and music in three signal-to-noise ratio (SNR) levels: 30 dB, 20 dB and 10 dB. The results show that ST strategy averagely improves the accuracy for music by 3.3% at 10 dB and 2.3% at 20 dB while keeping accuracy rate of speech almost unchanged. The speech classification rate is also lifted by 5.7% at 10 dB on average.