Some Results on Segmentation and Segment Classification in Continuous Speech

Abstract

A system for acoustic analysis of continuous speech is being developed as part of Lincoln Laboratory's effort in speech understanding. The acoustic system presently consists of two stages: the first produces a preliminary segmentation of the speech, as well as a gross classification of these segments into several acoustic categories; the second produces a finer classification intended to meet the needs of subsequent linguistic analysis. The current design and performance of the acoustic analysis system is outlined. This description is intended primarily as an interim report, since the system is being actively extended and refined. The first-stage segmentation and segment classification yields the following categories: vowel-like sound; volume dip within vowel-like sound; fricative-like sound; silence or voice bar; and burst. These categories are produced by a decision tree based upon energy measurements in selected frequency bands, derivatives, and ratios of these measurements, a voicing detector, and a few editing rules. Statistics will be presented describing the results of this analysis on a 31-sentence test corpus developed for an ARPA Segmentation Workshop held at Carnegie-Mellon University in July 1973. Second-stage classification includes a fricative identifier, which employs spectral measurements to group the fricative segments into phoneme-like categories. Results of this analysis for the test corpus will be presented. Discussion of additional work, including stop and vowel classification, will be given as time permits. [This work was sponsored by the Advanced Research Project Agency of the Department of Defense.]