A novel approach to speech-music discrimination based on rhythm (or beat) detection is introduced. Rhythmic pulses are detected by applying a long-term autocorrelation method on band-passed signals. This approach is combined with another, in which the features describe the energy peaks of the signal. The discriminator uses just three features that are computed from data directly taken from an MPEG-1 bitstream. The discriminator was tested on more than 3 hours of audio data. Average recognition rate is 97.7%.
展开▼
