Bayesian Nonnegative Harmonic-Temporal Factorization and Its Application to Multipitch Analysis

Since important musical features are mutually dependent, their relations should be analyzed simultaneously. Their Bayesian analysis is particularly important to reveal their statistical relation. As the first step for a unified music content analyzer, we focus on the harmonic and temporal structures of the wavelet spectrogram obtained from harmonic sounds. In this paper, we present a new Bayesian multipitch analyzer, called Bayesian nonnegative harmonic-temporal factorization (BNHTF). BNHTF models the harmonic and temporal structures separately based on Gaussian mixture model. The input signal is assumed to contain a finite number of harmonic sounds. Each harmonic sound is assumed to emit a large number of sound quanta over the time-log-frequency domain. The observation probability is expressed as the product of two Gaussian mixtures. The number of quanta is calculated in the -neighborhood of each grid point on the spectrogram. BNHTF integrates latent harmonic allocation (LHA) and nonnegative matrix factorization (NMF) to estimate both the observation probability and the number of quanta. The model is optimized by newly designed deterministic procedures with several approximations for the variational Bayesian inference. Results of experiments on multipitch estimation with 40 musical pieces showed that BNHTF outperforms the conventional method by 0.018 in terms of F-measure on average.