Speech and Singing Discrimination for Audio Data Indexing

This study investigates the technique of automatically discriminating speech from singing voices for audio data indexing. We propose a discrimination system based on both timbre and pitch feature analyses. In using timbre features, voice recordings are converted into Mel-Frequency Cepstral Coefficients and their first derivatives and then analyzed using Gaussian mixture models. In using pitch feature, we represent voice recordings as MIDI note sequences and then use bigram models to analyze the dynamic change information of the notes. Our experiments, conducted using a database including 600 test recordings from 10 subjects, show that the proposed system can achieve 94.3% accuracy.