Speech data base for testing components of speech understanding systems

Abstract

A set of over 700 sentences was designed to test acoustic effects due to various minimal contrasts in linguistic structures. One subset of sentences includes a comprehensive variety of phonetic sequences, to allow efficiently testing phonetic effects on prosodies, accuracy of phone recognition algorithms, and reliability of parameterization procedures such as ferment tracking. Another subset of sentences tests how acoustic correlates of stress and syntactic boundary locations (as determined by valleys in fundamental frequency contours) are altered as the position of stress within a constituent is moved. To study intonation patterns in isolation from the local fundamental frequency variations that accompany obstruents, many sentences have only sonorant sounds in them. Other sentences include only vowels and unvoiced consonants, to provide easy syllabication with energy measures. Sentences vary in complexity from simple declarative structures to sentences which introduce auxiliary verbs, prenominal adjectives, quantifiers, possessives, adverbs in various positions, there insertion, passives, yes/no and WH-questions, commands, and embedded structures such as restrictive and nonrestrictive relative clauses, complement structures, and conjuncts. The sentences were recorded by five male and three female talkers, with three repetitions spaced at least one week apart.