Speech spectrum analysis based on higher order crossings

In this paper, we propose a simple technique for extrapolating the spectra of speech signal based on real zero crossings and higher order crossings (HOC). In this approach, the speech signal is first subjected to a sequence of difference/sum filtering. Then counts of zero crossings are calculated over short overlapped frames to obtain a sequence of counts. This sequence represents the dominant frequency components of the speech signal over the analysis frames. The HOC based spectral characteristics are found to be distinct for each utterance. To evaluate the discrimination performance of the HOC counts, we consider them as feature vectors to train a speech recognition system. We found that the recognition rate achieved by using the HOC feature is comparable to what can be achieved by the commonly used mel frequency cepstral coefficients, which is more computationally expensive. This paper would pave the road for more extensive work on the HOC features and variants as applied to speech signals.