Phonetic design of overlapping speech in talk-in-interaction: A cross-linguistic study

Participants in a conversation often speak at the same time, leading to overlapping speech; indeed, it has been estimated that up to 13% of the speech we utter during normal conversation occurs simultaneously with that of another talker. Given that conversations are generally perceived to proceed smoothly, the high occurrence of overlap in conversation requires explanation. However, phonetic science has largely neglected overlapping talk, and its potential theoretical importance for our understanding of linguistic structure and function have hardly been explored. An understanding of overlapping talk is also important for a number of practical applications, including those in which human-computer interaction is achieved through a speech interface.

Conversations normally conform to a turn-taking model, in which participants wait for others to stop talking before talking themselves. However, overlapping speech can be used by conversational participants to compete for the turn - in other words, overlaps are initiated in order demonstrate that the overlapper is interested in taking over the turn immediately, not when the current speaker has finished. Conversations also contain a large number of non-competitive overlaps that have different conversational functions. A common example are so-called backchannel continuers such as 'uh-huh' that confirm the current talker's right to the turn.

Given that both competitive and non-competitive overlaps are well attested in conversation, the question arises as to what linguistic resources are employed by participants in order to display an overlap as turn competitive or as non-competitive. In answering this question, previous research has focused almost exclusively on the English language. There is therefore very little known about the similarities and differences between languages, with respect to overlapping talk, particularly in regard to its phonetic aspects. Also, previous work in the field has mainly used careful impressionistic listening rather than quantitative techniques.

Our proposal aims to address the limitations of previous work by conducting a cross-linguistic study of overlapping talk, combining quantitative with qualitative methods. We will record an audio/visual corpus of conversations held in two European languages - Standard Southern British English and Bosnian Serbo-Croatian. The conversational participants will be recorded in a naturalistic setting (such as a meeting room) in their home country. The corpus will allow us to determine whether the phonetic characteristics of overlapping talk are similar across different languages, and whether any differences relate to the particular system of accents and intonation used in the language. We will also use the corpus to determine whether overlapping talk that shares the same phonetic design in the two languages has a similar function.

The proposed project will also establish objective techniques for the analysis of overlapping talk in naturally occurring interactions. This will be achieved by exploiting recent developments in the field of speech technology relating to multi-channel recording and analysis of meetings, segregation of overlapping speech and pitch analysis of concurrent sounds. We anticipate that such techniques will allow us to extract phonetically relevant features from audio recordings of overlapping talkers in a naturalistic environment. A key issue, then, is whether analyses of overlapping talk based on objective acoustic analysis agree with previous analyses based on impressionistic listening.

Finally, we will make an annotated database of our audio/visual recordings available to other researchers via the Internet, and provide the facility for researchers to share their own annotations of the corpus. In this way, the corpus of recordings will be an evolving resource that will continue to benefit the research community beyond the life of the proposed two-year project.