Robust Hidden Markov Models for limited training data for birdsong phrase classification

Hidden Markov Models (HMMs) have been studied and used extensively in speech and birdsong recognition but they are not robust to limited training data and noise. This work present a novel method to training GMM-HMMs with extremely limited data—and possibly noisy—by sharing HMM components and generating more training samples that cover the variation of the models. We propose an efficient state-tying algorithm that takes advantage of unique characteristics of birdsongs. Specifically, the algorithm groups HMM states based on the spectral envelopes and fundamental frequencies, and the state parameters are estimated according to the group assignments. For noise-robustness, prominent time-frequency regions (time-frequency ranges expected to contain high energy for a particular HMM state) are used to compute the state emitting probability. In Cassin’s Vireo phrase classification using 75 phrase types, the results show that the proposed state-tying algorithm significantly outperforms both traditional state-tying ...