Automatic generation of prosody: comparing two superpositional systems

Abstract We face many options when designing a system thatautomatically generates prosody from linguistic andparalinguistic information. The literature provides severalcandidate phonetic models, phonological models and mappingtools to actually implement the system. We detail here somedimensions along which these models have to be compared. Weshow also that systems employing quite similar phonetic modelscan still have radically different approaches. We present resultsof a first evaluation comparing two systems using asuperpositional model of melody on a common multilingualprosodic database of spoken math formulae. We conclude thatprosodic models and intonation theories could certainly benefitfrom well-defined tasks and fair benchmarks. 1. Introduction It is a commonly accepted view that prosody crucially shapes thespeech signal in order to ease the decoding of linguistic andparalinguistic information by the listener.The present study compares two automatic prosodygeneration systems applied to a bilingual corpus. Our aim is notto designate a “winner” but to promote inter-system comparisonon common data as valuable aid to shed light on strengths andweaknesses of either system. The systems used here have to facedifferent challenges: the SFC is applied to German (its targetlanguage being originally French); the IGM has to deal both witha new type of document (spoken mathematical formulae) and anew language (French). The reader should have in mind that thechallenge for the latter demands more invasive adaptation of thesystem and results are to be considered as preliminary.After introducing the framework in which we apprehendprosody generation systems, the two systems are brieflydescribed (§3). Objective and subjective evaluation (§4) willenable us to detail some properties of the respective systems(§5).