Instrumental Sensibility of Vocal Detector Based on Spectral Features

Detecting voice in a mixture of sound sources remains a challenging task in MIR research. The musical content can be perceived in many different ways as instrumentation varies. We evaluate how instrumentation affects singing voice detection in pieces using a standard spectral feature (MFCC). We trained Random Forest models with song remixes for specific subsets of sound sources, and compare it to models trained with the original songs. We thus present a preliminary analysis of the classification accuracy results.