Automatic detection of head voice in sung musical signals via machine learning classification of time-varying partial intensities

The automatic detection of portions of a musical signal produced according to time-varying performance parameters is an important problem in musical signal processing. The present work attempts such a task: the algorithms presented seek to determine from a sung input signal which portions of the signal are sung using the head voice, also known as falsetto in the case of a male singer. In the authors’ prior work [Mysore et al., Asilomar Conf. Signal. Sys. Comp. (2006) (submitted)], a machine learning technique known as a support vector classifier [Boyd and Vandenberghe, 2004] was used to identify falsetto portions of a sung signal using the mel-frequency cepstral coefficients (MFCCs) of that signal (computed at a frame rate of 50 Hz). In the present work, the time-varying amplitudes of the first four harmonics, relative to the intensity of the fundamental, and as estimated by the quadratically interpolated fast Fourier transform (QIFFT) [Abe and Smith, ICASSP 2005], are used as a basis for classification. Preliminary experiments show a successful classification rate of over 95% for the QIFFT-based technique, compared to approximately 90% success with the prior MFCC-based approach. [Ryan J. Cassidy supported by the Natural Sciences and Engineering Research Council of Canada.]