INVESTIGATION OF STRESS PATTERNS FOR SPEECH SYNTHESIS BY RULE

In an earlier investigation, a procedure for synthesizing speech by rule was described. Although intelligible, the speech had an unnatural or machinelike quality, owing partly to the lack of suitable rules for controlling prosodic features of the utterance. The factors considered most likely to affect the prosodic features of synthetic speech were investigated experimentally. Several combinations of increments in duration and in fundamental frequency were tested using a paired comparison technique. The set of conditions producing the most natural sounding (i.e., least machinelike) synthetic speech was obtained using a team of six listeners and three test sentences. The parameter values corresponding to this set of conditions were used to synthesize a new group of five simple declarative sentences. These same five sentences were also synthesized using the earlier synthesis scheme. A second paired comparison experiment showed that the speech generated by the new scheme was preferred over the old in more than 80% of the comparisons. © 1969, Acoustical Society of America. All rights reserved.