Linear interpolation of spectrotemporal excitation pattern representations for automatic speech recognition in the presence of noise

This article is based on the study of new methods to improve recognition capabilities of automatic speech recognition in the presence of noise systems. Instead of trying to modify complex recognition models, the study is aimed at enhancing the input data´s reliability. This is achieved through processing of the acoustic representations of speech. One of these representations, called SpectroTemporal Excitation Pattern (STEP) is used in recognition systems with missing or unreliable data. One of the ideas behind this study was to increase the glimpsing areas in the STEP representations. And, because the glimpsing algorithm requires previous knowledge of the noise, another idea was to estimate noise characteristics, and base the glimpsing areas determination on these estimations. Preliminary tests were conducted with an HMM recognition system, but this will be the object of a future study.