Zero crossings as a means of obtaining spectral information in speech analysis

Zero-crossing patterns and spectral analysis using a filter bank are both a possible means in speech recognition for extracting information in the frequency domain that is useful in separating speech sounds. Results obtained from 60 speakers are given that show that the processing of segments of the 11 English vowels using a 13-channel filter bank give vowel separation comparable to that achieved with a 3-channel filter set followed by zero-crossing circuits. With prefiltering into 3 channels, zero-crossing patterns contain misleading and redundant information and it is shown that, by extracting only one zero-crossing distance during each pitch period, improved separation of low first formant vowels can be achieved with an economy in processed information. In vowels with closely spaced formants, both of which lie within the same analog filter band, the zero-crossing measure can give a mean of the two formant positions. The extent of spectral energy spread may be an important clue as well as its mean value, and a method is proposed for measuring it. In total, a much better understanding has been achieved of how filter-bank and zero-crossing analysis relate to each other and how best to extract zero-crossing information.