Another significant issue for the application of the LR framework is how the relevant
population should be sampled once it has been appropriately defined. In forensic DNA analysis, databases are collected using convenience sampling from blood banks and
disease screenings. This is possible for DNA since allele profiles are “uncorrelated with the means by which samples are chosen” (National Research Council 1996). However,
as highlighted by the multitude of sources of within-speaker variability in §2.2.5, speech variables are intrinsically affected by the situation in which they were elicited. This
means that it is extremely difficult to collect a sample of the relevant population which sufficiently matches the facts of the case at trial.
necessarily display some degree of mismatch with the facts of the case at trial, due to the wide range of factors affecting within- and between-speaker variation. The extent
to which such mismatch affects the resulting LR estimates is an empirical question and has received little attention in the literature. However, it is essential that the influence of
any mismatch is acknowledged and understood by practitioners in casework. There are currently three alternatives for assembling quantitative data for LR testing: case specific
data, existing non-forensic corpora and existing forensic databases. The benefits and limitations of these approaches are considered below. A further approach, of course,
which is not considered in detail here, is that the analyst estimates patterns in the relevant population based on experience and previous research.
2.4.1
Going and getting it(Rose 2007b)
Rose (2007b) argues that “we (forensic speech scientists) have . . . to be prepared to
go and get a suitable reference sample for each case.” The use of case-specific data provides the expert with much greater scope to control relevant elements of the facts of
the case at trial. An example of the collection of case-specific reference data in FVC is given in Rose (2013b). The limitations of the specific procedures in Rose (2013b) are
outlined in §2.2.3. There are also more general limitations of thegoing and getting it approach.
Firstly, there will still inevitably be some degree of mismatch with the facts of the case at trial, given that the expert is responsible for making subjective decisions over
which factors to control and which to ignore. In Rose (2013b), the limitations of the case-specific reference data were not considered in terms of their potential effect on
LR output. Secondly, there are considerable financial and time constraints imposed when conducting casework. Given these constraints, there is a danger that case-specific
reference data will have more shortcomings, in terms of the facts of the case, thanoff- the-shelf data (§2.4.2). Thirdly, such constraints mean that it is only possible to collect
reference data for analysing very short amounts of speech or a limited set of variables. For example in Rose (2013b) analysis is limited to the wordyesand the phrasenot too
bad. It is therefore considered prohibitively difficult to collect case-specific data for a componential analysis of a range of linguistic-phonetic variables (§1.1.3), especially
where variables occur in different utterances, words and phonological contexts.
2.4.2
Off-the-shelf
data
2.4.2.1 General corpora
It may be possible to use general corpora which were not originally collected for
forensic purposes in LR testing. A significant benefit of this approach is that data need not be collected for each case. This improves the time and cost efficiency of the analysis
and potentially extends the range of variables which may be analysed. The most suitable corpora are probably those collected as part of sociolinguistic research. Sociolinguistic
corpora often contain speakers controlled for numerous sociolinguistic factors, allowing for a definition of the logically relevant population with varying narrowness. The
breadth of sociolinguistic research means that data are available for a range of different regional and social groups. Such corpora also contain relatively long samples (c. 40-60
mins) and, in some cases, multiple samples of speakers in different speaking styles (e.g. spontaneous speech, ethnographic interview, read text). Examples of such corpora
include ONZE (Gordonet al. 2007), the Big Australian Speech Corpus (Wagneret al. 2010) and the Northern (British) Englishes corpus (Haddican 2008-2013).
However, the lack of forensic realism in such corpora is potentially problematic, since
they are likely to display considerable divergence from the facts of any case at trial. In particular, samples recorded for general corpora do not generally involve transmission
mismatch (Künzel 2001; Byrne and Foulkes 2004), mismatch in background noise and signal-to-noise ratio (SNR), or stylistic variability relevant for forensic purposes such
as speech under stress or different emotional states. Sociolinguistic corpora can also be very small, with few containing more than 30 speakers from the same sociolinguistic
community. Further, corpora containing multiple recordings from each speaker are generally made in a single session, rather than non-contemporaneously.
2.4.2.2 Forensic databases
Finally, there are a small number of databases available which were collected specifi-
cally for forensic purposes. The only existing forensic corpus for any variety of BrEng is the Dynamic Variability in Speech corpus (DyViS) (Nolan et al. 2005-2009; see
§3.1.1). Forensic databases also exist for AusEng (Rose 2007-2010; Morrisonet al. 2010-2013) and Standard Chinese (Zhang and Morrison 2011). In the field of ASR,
considerably more forensic databases are available (see Campbell and Reynolds 1999). Forensic databases have the benefit of having been controlled for the typical facts of
casework, such as transmission mismatch, non-contemporaneity samples and mismatch in speaking style. Such databases are also commonly much larger than general corpora,
allowing for testing using different subsets of the available data.
However, there are also limitations with forensic databases. There is limited availability
of forensic databases. In the case of BrEng, even DyViS is limited since it contains only speakers of Standard Southern British English (SSBE). This is inadequate for narrower
definitions of the logically relevant population, even with regard to regional background. Conversely, other forensic databases contain speakers from very wide, sociolinguisti-
cally heterogeneous populations, reflecting the Rose (2004) default relevant population based on sex and language. For example, Morrisonet al. (2010-2013) contains male
speakers of AusEng, with no control over other potentially sociolinguistically relevant factors. The relative lack of usable forensic databases is highlighted by French and
Harrison as a primary reason “for precluding the quantitative application of (the LR) approach” (2007: 142) in casework. Further, forensic databases, in particular those
used for ASR, often contain relatively small short samples for each speaker and little spontaneous material.