Document Type

Theses, Ph.D

Rights

This item is available under a Creative Commons License for non-commercial use only

Disciplines

Information Science

Publication Details

Successfully submitted for the award of Doctor of Philosophy (Ph.D) to the Dublin Institute of Technology, 2010.

Abstract

Inter-speaker accommodation is a well-known property of human speech and human interaction in general. Broadly it refers to the behavioural patterns of two (or more) interactants and the effect of the (verbal and non-verbal) behaviour of each to that of the other(s). Implementation of this
behavior in spoken dialogue systems is desirable as an improvement on the naturalness of humanmachine interaction. However, traditional qualitative descriptions of accommodation phenomena do not provide sufficient information for such an implementation. Therefore, a quantitative
description of inter-speaker accommodation is required. This thesis proposes a methodology of monitoring accommodation during a human or humancomputer dialogue, which utilizes a moving average filter over sequential frames for each speaker. These frames are time-aligned across the speakers, hence the name Time Aligned Moving Average (TAMA). Analysis of spontaneous human dialogue recordings by means of the TAMA methodology reveals ubiquitous accommodation of prosodic features (pitch, intensity and speech rate) across interlocutors, and allows for statistical (time series) modeling of the behaviour, in a way which is meaningful for implementation in spoken dialogue system (SDS) environments.
In addition, a novel dialogue representation is proposed that provides an additional point of view to that of TAMA in monitoring accommodation of temporal features (inter-speaker pause length and overlap frequency). This representation is a percentage turn distribution of individual speaker
contributions in a dialogue frame which circumvents strict attribution of speaker-turns, by considering both interlocutors as synchronously active. Both TAMA and turn distribution metrics indicate that correlation of average pause length and overlap frequency between speakers can be attributed to accommodation (a debated issue), and point to possible improvements in SDS “turntaking” behaviour. Although the findings of the prosodic and temporal analyses can directly inform SDS implementations, further work is required in order to describe inter-speaker accommodation sufficiently, as well as to develop an adequate testing platform for evaluating the magnitude of
perceived improvement in human-machine interaction. Therefore, this thesis constitutes a first step towards a convincingly useful implementation of accommodation in spoken dialogue systems.

Share

COinS