Voice on Trial: Forensic Speaker Comparison

In 1966, Lawrence Kersta became the first voice-identification expert to testify in court. He said his "voiceprints" were as reliable as fingerprints.

Durand Begault and Fausto Poza later traced the claim to a Nature study in which novice high-school girls compared isolated words from twelve speakers. Reported error rates were 0% to 3%. The task was closed-set: the corresponding sample was always present, participants knew that, and they could compare freely among exemplars. Those conditions differ substantially from casework in which the true speaker may not be represented. Kersta also manufactured spectrographs and profited from examiner certification. The term voiceprint encouraged an analogy with fingerprints, although voices vary with health, age, emotion, effort and transmission channel.

In 1978, the Second Circuit admitted spectrographic evidence in United States v Williams, relying on Tosi's 6.3% false-identification rate and analogies with handwriting and gun-barrel comparison. It described the central task as "the simple step of visual pattern-matching, a step easily comprehended and evaluated by a jury." Begault and Poza criticise that approach as subjective and impressionistic, noting the absence of evidence that trained examiners identify voices by ear more accurately than lay people. The expert, Lundgren, reported thirteen correspondences against a ten-point standard set by his trade association.

The court itself avoided the term voiceprint, writing that a spectrogram "has been often called a voiceprint. We avoid the term as potentially leading to an unwarranted association with fingerprint evidence." Current evidence should not be framed as visual voiceprint identification.

In 2009, Geoffrey Stewart Morrison argued in Science & Justice that forensic voice comparison was moving from categorical identification towards likelihood ratios, as DNA evidence had done in the 1990s. A likelihood ratio compares the probability of the observed voice evidence under a same-speaker proposition with its probability under a different-speaker proposition. An LR of 100 means the evidence is 100 times more probable under the same-speaker proposition. An LR of 1/1000 means it is 1000 times more probable under the different-speaker proposition. The numerator captures similarity and the denominator typicality; similarity in a common feature provides little support because it would also occur often among different speakers.

In a 2012 New South Wales fraud trial, a practitioner said there was "much to support a hypothesis that the defendant is the speaker," based on auditory assessment and spectrograms of the word "no." No performance figure for comparable conditions was given. Enzinger and Morrison built a statistical system using the same features, vowel-resonance movement and mean pitch, then tested it using a compressed mobile call and a reverberant police interview. Its Cllr was 0.834. A system that always returns "no information" scores 1, so those features provided little useful information under the simulated case conditions.

Kelly et al. tested the commercial VOCALISE system in 2019 using forensic_eval_01, a public case-like dataset containing 223 recordings from 61 speakers, with 111 same-speaker and 9720 different-speaker comparisons. Across six configurations, pooled Cllr ranged from 0.246 for a configuration adapted to case-like conditions to 0.462 for built-in configurations. Case-relevant adaptation nearly halved the cost.

The relevant question is the Cllr measured on validation data comparable in duration, channel, language and speaking style to the recordings in the case. A Tippett plot shows cumulative same-speaker and different-speaker comparisons against log LR and should include credible-interval bands. If performance is supported only by experience or training, the error rate may remain, in the words of the Angleton ruling quoted by Enzinger and Morrison, "unknown and may vary considerably, depending on the conditions of the particular application." Daubert asks whether the technique "can be (and has been) tested."

A likelihood ratio depends on several modelling decisions. The first is the relevant population. Its denominator asks how typical the questioned voice is among other speakers, so the analyst must define those speakers.

Hughes and Foulkes (2015) used the same New Zealand English vowel data with reference populations matched, mismatched or mixed by social class and age. For social class, the matched system had Cllr 0.513 and the mismatched system 0.659. For age, the corresponding values were 0.561 and 0.712. Changing between matched and mismatched populations shifted some individual LRs by more than a factor of ten. The authors note that "one cannot know for certain the sociolinguistic community to which the offender belongs." The chosen population must therefore be justified and its uncertainty acknowledged.

Be ready to identify the relevant population and explain its selection, describe calibration, and show that the validation data reproduce the case channel, codec, noise level and duration. Weakness in any of those components affects the meaning of the LR.

Two distinct human-factors issues arise in voice evidence: contextual bias affecting experts and memory limitations affecting earwitnesses.

Kukucka, Kassin, Zapf and Dror surveyed 403 forensic examiners in 21 countries in 2017. Participants estimated their own accuracy at a mean of 96%; 148, or 37%, claimed 100% accuracy. While 71% regarded cognitive bias as a concern in forensic science, 52% saw it as a concern in their own discipline and 26% thought their own judgements were affected. This is the bias blind spot. Seventy-one per cent believed they could reduce bias by trying to set expectations aside. The authors say effort is not a sufficient remedy because bias can operate automatically and without awareness. Procedural controls, including Linear Sequential Unmasking, regulate what information is received and when. Laboratories adopting them reported that they were neither onerous nor expensive.

Do not treat a calibrated system comparison and an earwitness identification as equivalent. They involve different mechanisms, performance evidence and safeguards. Expertise may improve technical comparison, but it does not establish immunity from contextual bias.

A voice case may now require authentication before speaker comparison. Liu and colleagues reported the ASVspoof 2021 challenge, which received submissions from 54 teams across three tasks, including fabricated speech "in the voice of a target speaker." On the deepfake progress data, 23 of 33 systems achieved equal error rates below 10%, with the best below 1%. On the less familiar evaluation data, every system exceeded 15%. Performance did not generalise from the development material.

A witness should not guarantee that a recording is a human utterance rather than a clone. A spoofing detector tested on known attacks may not generalise to new ones, while speaker comparison may assume that the questioned sample is authentic. Partial spoofing is especially difficult: the ASVspoof authors note that replacing "I won the election" with "I lost the election" changes the meaning while leaving most of the recording unaltered. The report should state which authentication assumptions were tested and which remain unresolved.