Acoustic feature mining for mixed speech and music playlist generation

The Internet and mobile phones allow customizing media content individually. In case of a radio program, beside a good selection of content, the quality of the transitions between pieces of audio material also play a significant role influencing the listening experience. This paper describes a study of speech to music transitions looking for patterns between the acoustic features and the subjective perception of the transition quality. In the course of the study a set of audio test data was created, a subjective opinion test for rating the quality of the transitions was conducted and acoustic features were extracted from both the pieces of speech and music. The collected data was analyzed using data mining methods. The most important pattern found in the data is that music and speech tempo, intensity range and Mel spectral coefficients make it possible to predict the quality of the match with a performance rate of 70%.