Improvement of voicing decisions by use of context

Voicing decisions in speech compression or recognition procedures are usually made in a context-free manner on successive fixed-length segments of speech. For a number of popular voicing statistics (zero-crossing rate, spectral slope, and low-frequency energy), the voiclng decision is improved by use of context, in fact by use of just the previous segment. For each statistic, instead of looking for a threshold that selects voiced segments, we use two thresholds, one if the last segment was called voiced and the other if the last segment was unvoiced. A typical improvement obtained by allowing this ´hysteresis´ in the voicing decision is a 15 percent drop in error rate.