As in Coleman and Walls (1974), assumptions about the alternative proposition may
be made based on factors which define sub-groups within the population at large, such as regional background, age and sex. This approach is referred to aslogical relevance
(Kaye 2004, 2008) and factors may be considered logically relevant if they affect the distribution of a variable in the wider population. This approach has been used
extensively in forensic DNA analysis. Since allele frequencies differ between racial groups (Gill and Clayton 2009), the logically relevant population is typically defined
by race. In the UK, three databases are used to evaluate DNA evidence based on broad racial groups: white Caucasian, Afro-Caribbean and Asian. As it is not possible to infer
racial background from the offender sample, multiple LRs are often presented based on different assumptions about the relevant population.
Variation in allele frequencies between sub-populations within racial groups has gener- ally been shown to be relatively minor (Gill and Clayton 2009; Baldinget al. 1996;
Budloweet al. 1999). Gillet al.(2000) assessed the level of regional variation in DNA profiles across 24 European populations making up the ENFSI DNA Short Tandem
Repeat (STR) Population Database.5 They concluded that for white Caucasians a single pan-European database is sufficient for generating stable LR output. Where such varia-
tion (e.g. regional variation due to high coancestry - regional groups displaying genetic similarity based on interrelatedness) is considered important, it may be accounted for
by incorporating acoancestry coefficient (FST) into the LR calculation (Balding and
Nichols 1994). Beyond race and regional background, the National Research Council
(NRC) states that, in some cases, it may also be necessary to consider other potentially logically relevant factors such as age and sex in forensic DNA analysis (1996: 30).
Applying the principles of logical relevance to FVC, Rose (2004: 4) claims that, in the
absence of a specific alternative proposition, the underlying assumption should be that
the voice in the offender sample does not belong to the suspect, but to “another same-sex speaker of the language.”Following this approach, the relevant population is defined by the sex and language of the offender. This definition has been used in almost all LR-
based FVC research (Kinoshita 2002; Alderman 2004a; Kinoshita 2005; Rose 2006;
Rose, Kinoshita, and Alderman 2006; Rose 2007a; Morrison and Kinoshita 2008; Morrison 2009b) and casework (Rose 2013b), and in the collection of FVC databases
(Rose 2007-2010; Morrisonet al. 2010-2013; Zhang and Morrison 2011).
2.3.1.1 Limitations
The Rose (2004) application of logical relevance makes two potentially problematic assumptions about FVC cases. Firstly, this approach assumes that language and sex
information are readily extractable from the offender sample. However, many cases present themselves where even these matters are not trivial (French et al. 2010: 145).
For example, Foulkes and French (2012) describe a case in which the unknown speaker on a telephone recording was assumed to be an adult female drug addict, but was in
reality a child. The issue of language is also complex due to issues of multilingualism, mobility and identity. Further issues are encountered which defining language more
narrowly in terms of regional dialect, since dialect does not equate directly to geo- graphical background. This is due to linguistic differences associated with the physical
and psychological spaces (Britain 2013), meaning that certain regional varieties are linguistically well-defined whilst for other dialects regional patterns may be much
more heterogeneous. Such incompatibility between social groupings and linguistic differences is reflective of the broader difficulties in defining what is meant by the term
speech community(see Patrick 2008).
Secondly, the Rose (2004) approach assumes that sex and language are the most
important sources of between-speaker variation, at least for those variables which are typically analysed in LR-based FVC (e.g. vowel formants). However, this reflects
a naïve view of the complexity of between-speaker variation in speech. Unlike in forensic DNA analysis, it is in principle possible for the sociolinguistically-informed
expert to determine considerably more demographic information about the offender, beyond sex and language (French and Harrison 2006). Furthermore, for many of the
available variables in auditory-acoustic FVC, sociolinguistic sources of variation other than language and sex may be far more relevant. For instance, there is no expectation
for marked differences between males and females in terms of VOT in British English (BrEng), but there may well be differences between ethnic groups (Heselwood and
McChrystal 2000).
However, only a limited number of studies have acknowledged the complexity of
between-speaker variation and the associated issues for defining the relevant population. Alderman (2004b) compared the Bernard (1970) and Cox (1999) databases of AusEng
as reference data using F1, F2 and F3 midpoints from the tense monophthongs /i a o 0 3/. LR testing was conducted using non-contemporaneous recordings of 11 speakers aged
between 18 and 26, and OLRs were calculated using naïve Bayes. Output was similar across the two sets although Cox (1999) (72.7%) marginally outperformed (1970)
(63.6%) by 9.1% in SS discrimination. Alderman concludes that both are useful for FVC, although “as more time passes and further change occurs (Bernard’s) usefulness
as a reference distribution will diminish” (2004b: 182). However, other sources of between-speaker variation, such as regional background and age, were not assessed.
Further, it is considered problematic to judge the usefulness of reference data purely on the output of speaker discrimination tests, rather than on whether it represents an
appropriate definition of the relevant population which answers the question asked by the court.
Roseet al.(2006) examined the speaker discriminatory value of AusEng /aI/ based on
a dual-target analysis (see §3.3.1) of the first three formants. As in Alderman (2004a), typicality was assessed using Bernard (1970) as reference data, and the issue of change
over time is again acknowledged. Based on a comparison with Cox’s (1999) data, Rose
et al. (2006) claim that the first target of F2 is now c. 100 Hz lower and that the second target of F1 is now c. 30 Hz higher. Despite acknowledging that such change is “impor-
tant” (p. 330), the potential effect on LRs was not investigated. Similarly, Morrison’s (2008) study of AusEng /aI/ acknowledges the use of heterogeneous reference data with
regard to regional variation and age (19 to 64 years). However, the logical relevance of these factors and their effect on the resulting LRs were overlooked. Only Zhanget al.’s
(2008) study of midpoint F1, F2 and F3 values for /i y/ in Standard Chinese extends Rose’s (2004) definition in controlling for age, sex and regional dialect.
The most extensive discussion of the complexity of the logically relevant population
is Loakes (2006), who investigated the performance of a test set of four pairs of male twins from Melbourne aged 18 to 20. Input consisted of F1, F2 and F3 midpoint
values from the eleven monophthongs of AusEng extracted from non-contemporaneous samples. The twin data were initially compared with the reference data from Sydney
from Bernard (1970). However, based on this pre-testing only a subset of the available formants from each phoneme were excluded from LR-based testing due to the levels of
divergence between the test and reference data. Loakes (2006: 214) offers a number of potential reasons for the divergence, including regional variation (test data = Melbourne,
reference data = Sydney), variation in the tasks performed by the test and reference speakers and the level of sociolinguistic heterogeneity (with regard to age, class etc.) in
the reference data. As suggested in Alderman (2004a) and Roseet al.(2006), processes of sound change in the time separating the test and reference data may also account for
the differences in formant frequencies. These factors lead Loakes (2006) to conclude that in defining the relevant population “tighter controls on (other) social variables might
also be applied” (p. 198) such as age, communities of practise, education, occupation and friendship groups.
However, there are also a number of problems with narrowly defining the relevant population according to sociolinguistic factors. The first relates to the appropriateness
of the expert defining the relevant population. Given that the relevant population is defined by the defence proposition, it is not, strictly speaking, the role of the expert to
make assumptions about it. However, there are good reasons to prefer decisions relating to FVC evidence to be made by the linguistics expert rather than by the court, legal
professionals or lay people (although this view is not universal; see §2.3.2). Secondly, an issue for the definition of logical relevance more generally is the paradox that without
knowing the identity of the offender, it is not possible to know for certain the logically relevant population of which he/she is a member. This applies equally to the general
Rose (2004) default as well as more specific propositions.