Automatic measurement and representation of prosodic features
Goangshiuan S. Ying, Leah H. Jamieson · Purdue e-Pubs (Purdue University System) · 1998
Effective measurement and representation of prosodic features of the acoustic signal for use in automatic speech recognition and understanding systems is the goal of this work. Prosodic features--stress, duration, and intonation--are variations of the acoustic signal whose domains are beyond the boundaries of each individual phonetic segment. Listeners perceive prosodic features through a complex combination of acoustic correlates such as intensity, duration, and fundamental frequency $(F\sb0).$ We have developed new tools to measure $F\sb0$ and intensity features. We apply a probabilistic global error correction routine to an Average Magnitude Difference Function (AMDF) pitch detector. A new short-term frequency-domain Teager energy algorithm is used to measure the energy of a speech signal. We have conducted a series of experiments performing lexical stress detection on words in continuous English speech from two speech corpora. We have experimented with two different approaches, a segment-based approach and a rhythm unit-based approach, in lexical stress detection. The first approach uses pattern recognition with energy- and duration-based measurements as features to build Bayesian classifiers to detect the stress level of a vowel segment. In the second approach we define rhythm unit and use only the $F\sb0$-based measurement and a scoring system to determine the stressed segment in the rhythm unit. A duration-based segmentation routine was developed to break polysyllabic words into rhythm units. The long-term goal of this work is to develop a system that can effectively detect the stress pattern for each word in continuous speech utterances. Stress information will be integrated as a constraint for pruning the word hypotheses in a word recognition system based on hidden Markov models.