Pitch determination of music signals using the generalized spectrum

This paper presents an algorithm for detecting and estimating pitch in acoustic audio signals using the Generalized Spectrum (GS). A performance evaluation of a GS-based and two classical (autocorrelation- and cepstrum-based) pitch determination algorithms was conducted on a set of wavetable-synthesized musical signals. The experiment separately evaluates the tasks of pitch detection and estimation. Pitch estimation performance is presented in terms of gross pitch errors (indicating algorithm stability) and mean-squared fine pitch error. Pitch detection performance is evaluated by a Receiver Operating Characteristic analysis of the detection statistics. Results demonstrate that the GS-based estimator generally performs worse than the autocorrelation and cepstrum-based methods. However, the GS-based method performed consistently better for the detection problem, especially at low signal-to-noise values.