A probabilistic analysis of the ratio spectrum

Recently, Lim and Harris (see IEEE International Symposium on Circuits and Systems, p.277-80, June 1998) developed a novel spectral representation called the ratio spectrum. The ratio spectrum is formed by taking the ratio of the power of a low-pass filtered signal to the power of the original unfiltered signal for all filter cutoff frequencies. Features extracted from the ratio spectrum have proven to be promising for phoneme recognition and speech compression. In this work we have shown that the ratio spectrum feature sets are competitive with standard feature sets used for automatic speech recognition. To understand the underlying properties of the ratio spectrum that contribute to the generation of a useful feature set, we derived the analytic expressions that describe the behavior of the ratio spectrum for a white noise model.