Speaker Comparison: Methods and the Expert's Role

The phonetic-auditory approach predates automated tools by decades and remains valuable in cases where audio quality is too poor for reliable acoustic measurement or where the linguistic content of the recordings matters as much as the acoustic signal. A forensic phonetician listens to both the questioned and reference recordings, attending to four main domains: voice quality (breathiness, creakiness, nasality, laryngeal settings), segmental features (how specific consonants and vowels are produced: the position of the tongue for /r/, the degree of aspiration on /t/, the colouring of vowel formants), prosodic features (habitual pitch range, speech rate, rhythm), and regional and social accent markers.

The auditory approach is trained and systematic, but it is not simple subjective impression. Practitioners work from established phonetic frameworks, and the conclusions they draw are supported by reference to specific time-aligned features in the recordings. The vulnerability of the approach is listener bias: if an examiner knows which recordings are 'supposed to' match, even unconsciously, the evaluation may be influenced. Blind evaluation protocols and independent second examiner review are the standard mitigation.

Acoustic-parametric analysis takes measurable quantities from the speech signal and subjects them to statistical comparison. The most widely used features in speaker comparison are formant frequencies, particularly F1 and F2 of vowels, which reflect the shape of the vocal tract and tend to be stable across sessions for a given speaker while showing consistent between-speaker differences. Fundamental frequency (F0, the pitch) statistics, long-term average spectra, and voice onset time are also used.

The Bayesian likelihood-ratio approach as formalised by Philip Rose (2002) and Geoffrey Morrison (2009) provides the statistical architecture. Formant measurements from the questioned and reference recordings are modelled as draws from speaker-specific distributions, and the LR is computed as the probability of observing those measurements from the same speaker divided by the probability from a different speaker drawn from a relevant population database. The choice of that population (which speakers represent 'different speakers'?) is consequential and must be justified in the report.

Modern automatic speaker recognition (ASR) systems used in forensic contexts are built on two main architectures. I-vectors, introduced by Dehak and colleagues around 2011 (Dehak at MIT CSAIL, Kenny and co-authors at CRIM, Montreal), represent a speech utterance as a fixed-length vector derived from a Gaussian mixture model trained on a large corpus of speech. The i-vector captures speaker-specific statistics in a low-dimensional space. Two i-vectors from the same speaker tend to be closer together in this space than two from different speakers.

X-vectors, proposed by Snyder and colleagues in 2018, replace the Gaussian mixture model with a time-delay neural network that learns discriminative features from thousands of hours of labelled training speech. X-vectors generally outperform i-vectors on standard evaluation benchmarks, particularly for short-duration speech. Both architectures use a PLDA (Probabilistic Linear Discriminant Analysis) back-end that converts the similarity score into an LR-calibrated value, accounting for within-speaker variability and intersession channel effects.

Speaker comparison is valid only when both the questioned and reference recordings adequately represent the speaker's natural voice under comparable conditions. Several factors degrade this comparability:

• Disguise: falsetto, whisper, pitch-shifting, accent imitation, and electronic voice changers can all alter formant patterns, fundamental frequency, and voice quality. Some disguises affect only one or two features; others are pervasive. The examiner must assess and document the degree of disguise and its effect on comparability.
• Stress and emotion: high stress raises fundamental frequency and can affect articulation clarity. Recordings of threatening calls or confrontational situations often show elevated F0 relative to neutral reference recordings. This reduces comparability and must be disclosed.
• Telephone and codec bandwidth: a narrow-band telephone recording (300-3400 Hz) removes the spectral information above 3400 Hz that some voice-quality and articulation features rely on. Formant comparison is still possible but some features are unavailable. Narrowband vs. wideband reference recordings should not be mixed without careful normalisation.
• Temporal mismatch: voice changes with age, illness, and deliberate vocal training. Reference recordings made years before the questioned recording may not represent the same vocal state. The examiner should note any significant temporal gap.

The IAFPA guidelines on forensic speaker comparison (Code of Practice, 2020) require that conclusions be expressed using a verbal scale tied to likelihood ratios, not as categorical assertions. The standard scale used by many English-language practitioners runs from 'strong support for the proposition that the voices come from different speakers' through 'inconclusive' to 'strong support for the proposition that the voices come from the same speaker', with intermediate gradations of 'moderate', 'limited', and 'very strong' support.

The IAFPA guidelines also specify that the report must state the methods used, the reference population chosen for LR computation, any factors that degraded the comparison (disguise, noise, bandwidth mismatch, short duration), and the qualifications of the examiner. Stating a conclusion without this context is considered inadequate.