Vowel Duration in an Artificial Talker Driven from a Phonetic Input

Abstract

At the 1961 Spring Meeting of the Society, we described a computer program which accepts for its input the names of phonemes punched on IBM cards and produces as its output the acoustic waveforms of connected speech. The program considers each phoneme to be composed of a context-free steady-state portion, together with initial and final transitions whose courses are determined by the identities of the phonemes on either side. Nominal steady-state and transition durations are stored for each phoneme, but, in actual practice, to produce natural phrasing the vowel durations have had to be adjusted by means of additional entries on the input cards. That situation has now been improved through a revision of the program whereby each vowel is stored in both stressed and unstressed forms, the desired version being signaled by a diacritic (/) in the input code. Remaining occasions necessitating the adjustment of stored times appear to reflect the workings of sentence stress.