Listener Preference and Comprehension Tests of Stress Algorithms for a Text-to-Phonetic Speech Synthesis Program.

摘要

Six different stress algorithms were tested to find a relatively simple set of rules that would work well with a text-to-phonetic program of limited size. Algorithms using spelling rules only (no lexical, syntactic, or semantic information) were compared with a monotone and with hand-placed stress (English pronunciation rules). In the preference test, an algorithm that uses statistical regularities in English to assign stress and timing was judged better than a monotone, random stress, and strictly alternating stress. Hand-placed stress was judged better than this algorithm only when combined with timing rules. Comprehension tests revealed differences only with unpracticed subjects, for whom the pattern of results was similar to that of the preference tests. (Author)