Spectral contrast normalization and other techniques for speech recognition in noise

Abstract

Three methods of improving speech recognition in noise are considered: energy thresholding, a noise-robust spectral representation called IMELDA, and a set of noise-robust spectral distortion measures. The spectral distortion measures can be seen as normalizing the contrast in the spectrum, a problem which can be transferred to the representation itself, making it computationally more efficient. In speaker-independent alphabet recognition tests in added steady white noise at various levels, IMELDA is shown to outperform a weighted cepstrum representation and be computationally more efficient. With this material and with digits recorded in trucks at a wide range of noise levels, performance is found to depend strongly on the threshold level. Contrast normalization is found to help, but only when the energy threshold is far from its optimum level.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>