Modeling State-Conditional Observation Distribution using Weighted Stereo Samples for Factorial Speech Processing Models. (arXiv:1503.02578v1 [cs.LG])

This paper investigates the role of factorial speech processing models in noise-robust automatic speech recognition tasks. Factorial models can embed non-stationary noise models using Markov chains as one of its source chain. The paper proposes a modeling scheme for modeling state-conditional observation distribution of factorial models based on weighted stereo samples. This scheme is an extension to previous single pass retraining for ideal model compensation and here we used it to construct ideal state-conditional observation distributions. Experiments of this paper over the set A of the Aurora 2 dataset shows that by considering noise models with multiple states, system performance can be improved especially in low SNR conditions up to 4% absolute word recognition performance. In addition to its power in accurate representation of state-conditional observation distribution, it has an important advantage over previous methods by providing the opportunity to independently select feature spaces for both source and corrupted features. This opens a new window for seeking better feature spaces appropriate for noise-robust tasks independent from clean speech feature space.

from cs.AI updates on arXiv.org http://ift.tt/185Z3y6

via IFTTT