Abstract

GRAVANO, Agustín. Mimicking the interlocutor's speech: Models of speaker entrainment for improving the nat uralness of spoken dialogue systems. Subj. procesos cogn. [online]. 2010, vol.14, n.2, pp.79-88. ISSN 1852-7310.

As speech processing technologies continue to improve, the old dream of creating a machine that talks gradually becomes real. The present interactive speech systems enable users to perform simple tasks such as banking transactions and hotel reservations, through verbal interaction. Despite being relatively successful, these human-computer conversations still have a long way to go regarding their naturalness: these systems tend to be described as "odd" or even "intimidating" by users. Among the main reasons for this lack of naturalness, is the flawed modeling of prosodic variation or the way some properties of speech (such as intonation, intensity and rhythm) change in verbal expressions. Current systems are still unable to handle these features correctly, both to understand the speech of the user as to produce synthesized responses. Prosodic variation is extremely complex in spontaneous speech, and it is well known that it´s affected by several levels of linguistic representation (lexical, syntactic, semantic and pragmatic). The present article focuses on a specific dimension of prosodic variation, known as "mimetization between interlocutors", which consists in the automatic alignment of speech features between the participants of a dialogue. After a general overview of the literature on these subjects, a research project in process that seeks to model the prosodic mimetizatin in dialogues is described.

Keywords : Speech processing; Dialogue; Prosody; Alignment.

        · abstract in Spanish     · text in Spanish     · Spanish ( pdf )