On frequency averaging for spectral analysis in speech recognition

Abstract

Many speech recognition systems use logarithmic filter-bank energies or a linear transformation of them to represent the speech signal. Usually, each of those energies is routinely computed as a weighted average of the periodogram samples that lie in the corresponding frequency band. In this work, we attempt to gain an insight into the statistical properties of the frequency-averaged periodogram (FAP) from which those energies are samples. Thus, we have shown that the FAP is statistically and asymptotically equivalent to a multiwindow estimator that arises from the Thomson's optimization approach and uses orthogonal sinusoids as windows. The FAP and other multiwindow estimators are tested in a speech recognition application, observing the influence of several design factors. Particularly, a technique that is computationally simple like the FAP's one, and which is equivalent to use multiple cosine windows, appears as an alternative to be taken into consideration.