An unsupervised approach to glottal inverse filtering

摘要

The extraction of the glottal volume velocity waveform from voiced speech is a well-known example of a sparse signal recovery problem. Prior approaches have mostly used well-engineered speech processing or convex L1-optimization methods to solve the inverse filtering problem. In this paper, we describe a novel approach to modeling the human vocal tract using an unsupervised dictionary learning framework. We make the assumption of an all-pole model of the vocal tract, and derive an L1 regularized least squares loss function for the all-pole approximation. To evaluate the quality of the extracted glottal volume velocity waveform, we conduct experiments on real-life speech datasets, which include vowels and multi-speaker phonetically balanced utterances. We find that the the unsupervised model learns meaningful dictionaries of vocal tracts, and the proposed data-driven unsupervised framework achieves a performance comparable to the IAIF (Iterative Adaptive Inverse Filtering) glottal flow extraction approach.