Some effects of speech materials on vocoder quality evaluations

Abstract

A set of six sentences was composed, of which the first contained only vowels and /r, w, y/. Besides vowels, the next three sentences contained only nasals, only fricatives, and only stops and affricates, respectively. The final two sentences contained between them almost all the consonants of English. The sentences were recorded by three male and three female talkers, and the recordings were passed through a set of 12 simulated LPC vocoders, whose bit rates were equalized at about 2600 bits/sec (nonoptimal coding). The processed sentences were used in a quality rating task (each talker x system x sentence presented once), and also in a rank‐ordering task, in which a subject ranked all systems for each talker x sentence combination. The same subjects served in both experiments. The responses were subjected to multidimensional scaling analysis. The results show that the relative quality of a pair of systems can be strongly affected both by the speech material and by the talker. Implications for quality eva...