SINGING VOICE CHARACTERIZATION FOR AUDIO INDEXING

摘要

To extract the content of audio documents, the first steprnin many approaches is to segment the signal in primary components,rnsuch as music and speech. Very few attention hasrnbeen brought to the detection of the singing voice.rnIn this paper, we propose simple parameters (vibrato andrnharmonic coefficient) and an original segmentation based onrna sinusoidal segmentation to characterize the singing voice.rnThis information is then mixed with those issued from arnspeech/music decomposition.rnWe test this classification system on a database composedrnof various types of sound. We first test our system in a classificationrntask, then in a detection task. In both cases, thernresults are good. In our classification system, the only misclassificationsrnare due to very rare musical styles. In the detectionrntask, our system misses some of the singing voice segments,rnbut we observe very few false-alarm.