Factorial Speech Processing Models for Noise-Robust Automatic Speech Recognition

This paper presents an introduction of factorial speech processing models for noise-robust automatic speech processing tasks. Factorial models try to use more noise information rather than other robustness techniques for better generative modeling of speech and noise and the way they are combine together. Since factorial models were not completely successful in noise-robust speech processing applications while they have significant achievements in other speech processing areas in the past, we decide to reconsider them and evaluate their effects in the Aurora 2 task. In addition to Aurora noises, two more regular noises are examined in our experiments including Helicopter and Locomotive engine noises. Experiments show that these models are successful when we faced with destructive noises in addition to their unexpected improvements for non-regular non-stationary noises like Babble.