Singer-Dependent Falsetto Detection for Live Vocal Processing Based on Support Vector Classification

We present and analyze a machine learning technique to determine from an input sung vocal waveform if falsetto (also known as the head voice) is being used. Such a system may be used to tune signal processing parameters, ideally in real-time, for such applications as intelligibility enhancement of high-pitched sung notes, and other musical systems which tune signal processing parameters according to detected performance parameters. Our falsetto detector uses a support vector classifier trained on mel-frequency cepstral coefficients (MFCCs) computed from a newly collected database of anechoic sung notes. It is shown to give correct classification with better than 95% accuracy.