Speech/music discrimination based on spectral peak analysis and multi-layer perceptron

This study presents a new Speech/Music discrimination method based on spectral peak feature and Multilayer Perceptron. The focus was on feature extraction that reflects spectral peak duration characteristics and high performance using small number of train dataset. Spectral peak features were extracted from audio spectral peak tracks and the feature was normalized by length of segment. Then, we grouping the frequency channel to reflect the spectral distribution. For train, only 25 seconds of speech (Korean) and 50 seconds of music are used This method was evaluated on speech and music for 24,258 seconds of audio data. An average accuracy was 96.58% for speech and 91.82% for music. The results of this experiment found that proposed method was suitable for Speech/Music discrimination.