Regression-based Melody Estimation with Uncertainty Quantification (2505.05156v1)

Abstract: Existing machine learning models approach the task of melody estimation from polyphonic audio as a classification problem by discretizing the pitch values, which results in the loss of finer frequency variations present in the melody. To better capture these variations, we propose to approach this task as a regression problem. Apart from predicting only the pitch for a particular region in the audio, we also predict its uncertainty to enhance the trustworthiness of the model. To perform regression-based melody estimation, we propose three different methods that use histogram representation to model the pitch values. Such a representation requires the support range of the histogram to be continuous. The first two methods address the abrupt discontinuity between unvoiced and voiced frequency ranges by mapping them to a continuous range. The third method reformulates melody estimation as a fully Bayesian task, modeling voicing detection as a classification problem, and voiced pitch estimation as a regression problem. Additionally, we introduce a novel method to estimate the uncertainty from the histogram representation that correlates well with the deviation of the mean of the predicted distribution from the ground truth. Experimental results demonstrate that reformulating melody estimation as a regression problem significantly improves the performance over classification-based approaches. Comparing the proposed methods with a state-of-the-art regression model, it is observed that the Bayesian method performs the best at estimating both the melody and its associated uncertainty.