INDIVIDUAL AND GROUP VARIATION IN DISFLUENCY FEATURES: A CROSS-ACCENT INVESTIGATION Kirsty McDougall, 1 Martin Duckworth 2 and Toby Hudson 1 1 Phonetics Laboratory, Department of Theoretical and Applied Linguistics, University of Cambridge; 2 Duckworth Consultancy Ltd kem37@cam.ac.uk, msd@duckworth-consultancy.co.uk, toh22@cam.ac.uk ABSTRACT A study of individual differences in the fluency disruptions of speakers of two different accents, Standard Southern British English (SSBE) and York English is presented. Distributions of rates of occurrence per 100 syllables are examined for filled and silent pauses, repetitions, prolongations and (self-)interruptions, and subcategories of these. Patterns of occurrence of disfluency features show considerable between-speaker variation in both SSBE and York English. Similar ranges of speakers’ overall disfluency rates are exhibited by both accents, but cross-accent differences are present in the patterning of some disfluency feature categories. The results suggest that a detailed record of disfluency features is a useful additional tool in forensic speaker comparison. Keywords: disfluency, individual differences, accent differences, forensic speaker comparison 1. INTRODUCTION Features of speech related to fluency such as silent and filled pauses, sound prolongations, repetitions and self-interruptions exhibit variation between speakers. However, the range of individual variation in non-fluency phenomena in speakers identified as having normally fluent speech has received little attention in phonetic research. Non-fluency pheno- mena have clear potential for individual variation since they are likely to play a part in the planning of speech (e.g. [3]), may be influenced by psycho- or socio-linguistic demands, and are difficult – maybe even impossible – to control consciously. Analysis of individuals’ use of disfluencies has great potential for application in forensic speaker comparison cases, in which voice recordings of an unknown speaker committing a crime and a suspect are compared. The bulk of the literature on speaker- distinguishing properties of speech for forensic applications has focussed on phonetic features which bear a direct relationship with a speaker’s anatomy, e.g. mean f0 and formant frequencies (e.g. [3]). Inv- estigating the speaker-distinguishing potential of dis- fluency features focusses on a very different aspect of a speakers’ performance: speech features which are behavioural rather than anatomical. Further, disfluency features are largely realised through the temporal domain and therefore generally well- preserved in the poor recording conditions of forensic cases where telephone transmission and background noise are typical. This is in contrast to the ‘anatomical’ features mentioned above which are conveyed through spectral information for which adverse recording conditions are more problematic. Developing an understanding of individual variation in normal non-fluency behaviour is also of importance in speech and language therapy where knowledge of the extent of disfluency behaviours among speakers with no speech or hearing problems provides a reference point for therapists working with stammering speakers. Although the study by Roberts et al. [7] on non-stuttering adults found relatively few disfluencies in their sample, the analysts counted some phenomena only if they sounded like stuttering. Eklund [1] on the other hand used a different metric and found that non-stuttering speakers produced a larger number of disfluencies. The taxonomy and methods of counting disfluencies require careful definition. [11] presents a study of individual behaviour in the fluency disruptions of 20 male SSBE speakers from the DyViS database [6] undertaking a simulated police interview and a telephone call with a ‘friend’. The rate of occurrence per 100 syllables of a range of disfluency features was calculated for each speaker in the two speaking styles. Results showed that individuals varied considerably in their rates of fluency disruptions and that individual differences were present in their ‘disfluency profiles’, i.e. the types of disfluencies each speaker produced. Disfluency features also showed a degree of within- speaker consistency across the two speaking styles. The present study investigates whether patterns of disfluency differ across different accent/cultural backgrounds, by comparing the disfluency behaviour of speakers of SSBE and of York English. The study considers whether the range of individual variation observed in SSBE mirrors that observed in York English, whether disfluency features could be considered accent-independent to any extent, and the implications of these findings for forensic phonetics.