Using zero crossing counts to provide discriminative information in isolated word recognition
Alexander I. Rudnicky, Alexander H. Waibel, Neeraja V. Krishnan · The Journal of the Acoustical Society of America · 1981
Template matching systems that use a purely spectral representation are potentially insensitive to some important phonetic dimensions of speech. For example, [PI] is often confused with both [BI] and [TI] because the necessary vocalic—aspirated—fricative discrimination cannot be reliably made on the basis of spectrum alone. The zero crossing count is known to contain sufficient information for this and for other discriminations. We evaluated the usefulness of the zero crossing count in a set of experiments that used a highly confusable subset of the alpha-digit vocabulary (3, b, c, d, e, g, p, t, v, z). In these experiments, a zero crossing count was added to a 15 coefficient spectral representation and the resulting vector was used to calculate frame distances within a dynamic warping algorithm. Testing over a corpus of 800 utterances (eight talkers), we achieved a 10% reduction in error rate. In order to obtain optimal improvement, the count had to be clipped to a restricted range, quantized to a small number of levels, and weighed appropriately relative to the spectral information. Further investigation determined that the addition of zero crossing information aided recognition by improving the distinctiveness of confusable utterances and not by improving the dynamic warping alignment of the spectral information.