Instrumentation & Measurement Magazine 24-1 - 39

emotions, etc.), and another that takes into account dimensions
that exist regardless of the presence of interaction (e.g., language or dialect, social context, etc.). In this regard, it is possible
to understand how speech analyses are (still) extremely difficult to generalize, if not at the cost of delimiting the observed
sample. This is not possible in forensics, where the recordings
are provided by third parties, and cannot be defined a priori.
Moreover, the prosodic behavior is reflected on the phonetic
level, and precisely on formants and other acoustic features
used in investigations, so that introducing complementary and
statistical analysis therefore becomes necessary.
The issue of variability at the phonetic level has been recently fully investigated by a new field of linguistic research
known as sociophonetics. By combining the methodological
and theoretical considerations of previous sociolinguistics, sociophonetics has focused on the variation of acoustic features
in the production of speech as a means to signal not only the
speaker's identity but also his/her orientation towards the
hearer or the emotions attached to a certain topic. Thus, when
explored from a sociophonetic perspective, the variability is
not intended as a simple noise in the data, but as a resource
used by speakers in everyday interactions, most of the time
without being aware of it. In this respect, phonetic cues provide social meaning.
One of the principal phonetic variables used in sociophonetic research, as well as in previous sociolinguistic and
dialectological inquiries, are vowels. In particular, measures of
vowel quality (including the vowel's height, advancement and
lip rounding) and duration have been approached differently
and with various tools. Vowels' formants remain the main
clues for analyzing vowels' variation. The first formant (F1) refers to the tongue's height along a high-mid-low continuum,
and presents in an inversely proportional way: that is, vowel
[a] is pronounced with the lowest tongue height and shows
the highest F1 values. The second formant (F2) refers to the advancement of the tongue, by distinguishing between front and
back vowels (e.g., [i] vs. [u]) with the more front vowels having
the highest F2 values. Lip rounding is usually associated with
a variation in both F2 and F3 (i.e., the third formant) values,
and this is particularly helpful for the study of those languages
(e.g., English) in which the opposition between rounded and
unrounded vowels is also phonological (i.e., it is a means for
distinguishing between different words). As a side note, it is
possible to highlight that phoneticians used to transcribe between square brackets speech sounds as they were actually
pronounced by the speakers. Phonetic notation could be very
precise and refers to a set of characters and symbols collected
in the International Phonetic Alphabet (IPA, cf. www.internationalphoneticalphabet.org). It should be noted, however,
that a very accurate phonetic notation is not easily readable by
non-experts, therefore being quite unusable outside its specific
field of research (e.g., for forensic purposes).

Intra- and Inter-speaker Variability
Since formants' values correspond to specific frequencies
in the sound signal, they are highly variable according to
February 2021

speakers' specific physical attributes: for instance, women
and children notoriously have higher formants' values with
respect to men. This means that speaker's specific characteristics, including both biological and social features, play a role
in shaping formants' variability in speech, also across different
speaking situations (for an overview, cf. [10]). A way to solve
this inter-speaker problem is through formants' normalization with various formulas (e.g., by transforming hertz values
in bark). However, the technician must be warned against the
use of normalization procedures, especially in small speech
samples, because they tend to squeeze variability, thus making
it unusable for sociophonetic analysis. A recent work has also
demonstrated how bark normalization drastically reduces
within-speaker's semi-automatic recognition in a possible forensic setting [11].
Furthermore, vowels' formants could be measured with
different techniques and at different points in time. A main difference is between static and dynamic approaches to formants'
measurements: the first centers on formants' values as extracted at vowels' midpoints, whereas the latter focuses on the
variation of F1 and F2 values through 5 to 7 timepoints during
vowels' durations [10].
Another perspective considers not only the single formant's variation, but the whole vowel space as created by all
the vowels pronounced by a speaker (or, at least, the cardinal
ones). Vowel space could be measured through Euclidean distances but also by using F1-F2 co-variation as cardinal points
on a Cartesian plane, mapping the movements of vowels in a
bi-dimensional space. As we will see, there are tools that help
in the visual representation of this variation, thus being a resource also for forensic purposes.

Signal Variability
The variability of the signal (and consequently of the phonetic
parameters) is also influenced by the quality and context of the
recording, which in the forensic field can change a lot according to the environment of the recording that can range from
a crowded restaurant to a silent street. Moreover, the analysis of vowels' formants variation across recording modalities
shows that formants are also modified by the compression format [11].
Comparing the voices of possibly the same speaker from
different recording devices for forensic purposes is possible
[12], but a qualitative analysis must be combined with a quantitative one.

Qualitative and Quantitative Analysis
In forensic settings, reliability of the results is particularly relevant because of the possible juridical consequences of the
reports provided by different experts. In voice comparison, the
most accredited method for statistical elaboration of formants'
values is based on a Bayes factor, obtained as a likelihood ratio of the marginal likelihood of two competing hypotheses
(usually a null and an alternative) in order to obtain a discrimination score. However, it requires extrapolation of a high
number of vowels according to their type, and this is often not

IEEE Instrumentation & Measurement Magazine 39

http://www.internationalphoneticalphabet.org http://www.internationalphoneticalphabet.org

Instrumentation & Measurement Magazine 24-1

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-1