The Use and Utility of Localised Speech Forms in Determining Identity Corpus - Vowel Formant Frequency Data, 2016-2019

Llamas, Carmen and French, Peter and Watt, Dominic (2021). The Use and Utility of Localised Speech Forms in Determining Identity Corpus - Vowel Formant Frequency Data, 2016-2019. [Data Collection]. Colchester, Essex: UK Data Service. 10.5255/UKDA-SN-853956

The project aims to investigate variation in accents of English across Northeast England. We focus on the speech of working-class people from Newcastle, Sunderland and Middlesbrough. It is well established that these accents differ significantly from one another, but it is less clear what particular pronunciations lead listeners to group the varieties as one accent (typically, 'Geordie'). The project seeks to identify specific features that cause the accents to be classed together, such that they are heard to be distinct from other accents of northern Britain. Simultaneously, we will ascertain what sounds listeners use to classify speakers into subgroups, e.g. from Newcastle versus Sunderland. The role of socio-economic class is of key importance here. The localised patterns that allow listeners to assign speakers to one group or another are, according to the literature, traditionally associated with speakers from lower socio-economic groups, i.e. the 'working class'. However, in this project we ask whether clear differences exist between the speech of working-class speakers who routinely travel for work or leisure and that of economically marginal individuals we might label 'never worked/long-term unemployed'. Plentiful sociological research on this social divide exists, but how it impacts on people's speech has not yet been investigated systematically. We will also take gender and age differences into account, and assess the data for signs of sound change. It has been shown that greater mobility promotes linguistic uniformity through convergence of speech habits, while limited mobility has the opposite effect. We therefore predict that more mobile members of our sample would converge linguistically over time and across the three localities, while the economically marginal groups in each place would become more divergent. The project's second strand concerns speech variation at the level of the individual rather than the group or community. This is particularly relevant in the forensic domain, wherein individual identity is crucial. In criminal investigations, forensic speech analysts perform two main tasks. The first, speaker profiling, involves attempting to specify the geographical and social origins of an unknown speaker from a recorded sample of his/her speech, so as to assist the police in identifying potential suspects. The task requires detailed, up-to-date information about the speech of the community/ies to which the unknown speaker may belong. Given its focus on the identification of highly localised speech forms, the corpus produced by the proposed project will satisfy those requirements. The second forensic task is speaker comparison. Here, the expert compares two speech samples, and assesses the likelihood that they were produced by the same or different speakers. Increasingly, this is done in an automated way, using software that extracts information about the acoustic properties of the recordings. The level of similarity between the two samples is evaluated in the context of a relevant 'background population' of recordings of speakers with the same or similar accents. This yields a measure of the samples' typicality. The problem with this approach is that ideally one ought to collect a new corpus for every case, which is likely to be prohibitively expensive. It would be advantageous, therefore, if acoustic parameters in the speech signal which are relatively insensitive to accent variation could be identified. The proposed project tests whether this is possible by combining the Northeast recordings with those from an existing corpus of a markedly different accent (Standard Southern British English). If the approach proves legitimate, its practical value to the forensic speech analysis community would be considerable. Findings emerging from this two-stranded project will therefore benefit the relevant academic communities as well as having significant applied utility in the field of forensic speech science.

Data description (abstract)

The data deposited are taken from fieldwork recordings undertaken with speakers from the three fieldwork sites, Newcastle, Sunderland and Middlesbrough in the Northeast of England. From each locality, 40 informants were recorded, giving a total of 120 informants. The key data in the file All_formants_July_2018.xlsx are vowel formant frequency measurements, in Hertz, for the peripheral monophthongs denoted by the lexical set keywords FLEECE, TRAP, START/PALM, BATH, DRESS, and NORTH/FORCE (after Wells 1982). In the varieties of English represented in the TUULS corpus, SQUARE is often a monophthong (represented here by [e:]) rather than an ingliding diphthong. We have also included formant measurements for the relevant vowels of the commA and lettER sets, plus tokens of schwa which are not exemplars of commA or lettER. This file is the master file containing the aggregated data for 120 individual participants (denoted P then participant number). There are 16,390 rows in the spreadsheet. The data are comprised of metadata about the participant and the group to which s/he belongs, the software settings, and characteristics of the recording from which each observation was drawn. We also include data taken from a subsample of 60 male speakers from the TUULS corpus. The data in the file IJSLLheatmapdatawithspeakers.csv are ‘speaker scores’ for the 60 male speakers from the TUULS corpus, and 100 male speakers from the DyViS (Dynamic Variability in Speech) corpus (Nolan, McDougall, de Jong & Hudson 2009). A speaker score is a measure of the difference between a pair of statistical models each representing the resonance characteristics of an individual speaker’s voice. The first 100 rows in the spreadsheet represent the DyViS speakers, who are labelled 001’ to 121’, as there are gaps in the sequence of number labels. The second block of 60 rows contains the data for the TUULS speakers (labelled using a ‘P’ prefix followed by a number). The columns are arranged according to the same sequence.

Downloads

Altmetric