Forensic Speaker Identification

Phil Rose

• Publisher: CRC Press
• Year: 2002
• eBook ISBN: 9781134486182
• Topic: Law
• Subtopic: Forensic Science
• Index: Law

A voice is much more than just a string of words. Voices, unlike fingerprints, are inherently complex. They signal a great deal of information in addition to the intended message: the speakers' sex, for example, or their emotional state, or age. Although evidence from DNA analysis grabs the headlines, DNA can't talk. It can't be recorded planning,

1 Introduction

Paul Prinzivalli, an air freight cargo handler in Los Angeles, stood trial for telephoning bomb threats to his employer, Pan Am. He was suspected because he was known to be a disgruntled employee, and because some Pan Am executives thought that the offender’s voice sounded like his. Defence was able to demonstrate with the help of forensic-phonetic analysis that the offender’s voice samples contained features typical of a New England accent, whereas Prinzivalli’s accent was unmistakably from New York. To untrained West Coast ears, the differences between New York and Boston accents are not very salient; to ears trained in linguistic and phonetic analysis, the recordings contained shibboleths galore. Reasonable doubt was established and Prinzivalli was acquitted (Labov and Harris 1994: 287–99).

In a case of the kidnapping and murder of an 11-year-old German girl (Künzel 1987: 5, 6), considerable agreement was found between the voice samples of a suspect and that of the kidnapper and, on the basis of this and other evidence, the suspect was arrested. Subsequently, a more intensive comparison between offender and suspect voice samples yielded yet more similarities. The man confessed during his trial.

In another American case involving a telephoned bomb threat (Hollien 1990: 51), the defendant had been identified by his voice. However, it was clear to forensic phoneticians even from an auditory comparison that the voices of the defendant and the offender were very different. For example, the offender’s voice had features typical of someone who spoke English as a second language. The case was dismissed.

In the late 1990s in Australia, in a case concerning illegal drug trafficking (Q. vs. Duncan Lam 1999), police intercepted 15 incriminating telephone conversations containing 31 voice samples in Cantonese. Forensic-phonetic analysis was able to assign these 31 voice samples to three different speakers. Since the police, but not the analyst, knew the identity of some of the samples, not only could two of the speakers be identified, but the accuracy of the identification could also be checked.

Another recent Australian case involved the interception of telephone conversations between two brothers, one of whom was charged with drug-related matters. Defence claimed that their voices were so similar that the incriminating recordings could not be attributed to the suspect. A forensic-phonetic analysis was able to show that the brothers’ voices were distinguishable. Although their voices were indeed acoustically very similar in many respects, they still differed in others, and in particular they both had different ways of saying their ‘r’ sound.

Forensic speaker identification

Expert opinion is increasingly being sought in the legal process as to whether two or more recordings of speech are from the same speaker. This is usually termed forensic speaker identification, or forensic speaker recognition. As the examples above show – and many more could be cited – forensic speaker identification can be very effective, contributing to both conviction and elimination of suspects. Equally importantly, the examples also demonstrate the necessity for expert evaluation of voice samples, since three of them show how the truth actually ran counter to the belief of naive listeners.

Forensic speaker identification of the type described in this book – that is, using a combination of auditory and acoustic methods – has been around for quite a long time. Germany’s Bundeskriminalamt was one of the first institutions to implement it, in 1980 (Künzel 1995: 79), and the first conference on forensic applications of phonetics was held in the United Kingdom in 1989. Yet there is still a considerable lack of understanding on the part of law enforcement agencies, legal practitioners, and indeed phoneticians and linguists, as to what it involves, what constitutes appropriate methodology, what it can achieve, and what its limitations are. In a 1995 Australian Court of Appeal report, for example, (Hayne and Crockett 1995: 2, 3) some perceived ‘weaknesses in the science’ [of speaker identification] were explicitly listed, some of which were inaccurate (Rose 1996d). The aim of this book is to make explicit and explain in detail for the relevant professionals what is involved in forensic speaker identification (FSI), and to clarify the problems of inferring identity from speech under the very much less than ideal conditions typical in forensics.

Forensic phonetics

Forensic speaker identification is a part of forensic phonetics. Forensic phonetics is in turn an application of the subject of phonetics. Different experts have slightly differing opinions on the exact subject matter of phonetics, and even whether it constitutes a discipline (Kohler 2000; Laver 2000; Ohala 2000). However, the following characterisation will not be controversial. Phonetics is concerned primarily with speech: it studies especially how people speak, how the speech is transmitted acoustically, and how it is perceived.

As well as forensic speaker identification, forensic phonetics includes areas such as speaker profiling (in the absence of a suspect, saying something about the regional or socioeconomic accent of the offender’s voice(s) ); the construction of voice line-ups; content identification (determining what was said when recordings are of bad quality, or when the voice is pathological or has a foreign accent); and tape authentication (determining whether a tape has been tampered with) (French 1994: 170; 182–4; Nolan 1997: 746). This book will not be concerned with the latter three areas of voice line-ups, content identification and tape authentication, and only very indirectly with speaker profiling.

Readership

It is a good idea to be specific early on about the intended audience. I have had several types of readers in mind when writing this book. As just mentioned, I have written it to help members of the legal profession, the judiciary, and law enforcement agencies understand what forensic speaker identification is about. This will help them when requesting a forensic-phonetic investigation and help them understand and evaluate forensic-phonetic reports and evidence. I have also thought of linguistic phoneticians, budding and otherwise, who might be lured away from describing how languages differ phonetically to do valuable forensic-phonetic research into the complementary area of how their speakers differ. This book is also intended to be useful to those contemplating a professional career in forensic phonetics. I hope that the speech science community at large, including those working in the area of automatic speaker identification, will find much that is of use in the book. I hope, too, that the book will be accessible to both graduate and advanced undergraduate students in all related disciplines. Finally, I hope that the book will help stimulate the interest of statisticians who are thinking of researching, or supervising, forensic-statistical topics: there is especially much to be done in this area!

The take-home messages

The most important points that the book will make are these:

• The forensic comparison of voice samples is extremely complex.
• In the vast majority of cases the proper way to evaluate forensic speech samples, and thus to evaluate the weight of the forensic-phonetic evidence, is by estimating the probability of observing the differences between them assuming that the same speaker is involved; and the probability of observing the differences between them assuming that different speakers are involved. This method is thus inherently probabilistic, and as such will not yield an absolute identification or exclusion of the suspect.
• The two main problems in evaluating differences between samples are (1) differential variation in voices, both from the same speaker and from different speakers, and (2) the variable and generally poor degree of control over forensic speech samples.
• Speech samples need to be compared both acoustically and auditorily, and they also need to be compared from the point of view of their linguistic and nonlinguistic features.
• Forensic speaker identification requires expert knowledge of not just one, but several different specialist areas related to speech-science. These include sub-areas of linguistics, acoustics and statistics. This is in addition to the knowledge of how to interpret the results forensically.

Argument and structure of the book

Speaker identification in the forensic context is usually about comparing voices. Probably the most common task involves the comparison of one or more samples of an offender’s voice with one or more samples of a suspect’s voice. Voices are important things for humans. They are the medium through which we do a lot of communicating with the outside world: our ideas, of course, but also our emotions and our personality:

The voice is the very emblem of the speaker, indelibly woven into the fabric of speech. In this sense, each of our utterances of spoken language carries not only its own message, but through accent, tone of voice and habitual voice quality it is at the same time an audible declaration of our membership of particular social regional groups, of our individual physical and psychological identity, and of our momentary mood.

Laver (1994: 2)

Voices are also one of the media through which we (successfully, most of the time) recognise other humans who are important to us – members of our family, media personalities, our friends and enemies. Although evidence from DNA analysis is potentially vastly more eloquent in its power than evidence from voices, DNA can’t talk. It can’t be recorded planning, carrying out or confessing to a crime. It can’t be so apparently directly incriminating. Perhaps it is these features that contribute to the interest and importance of FSI.

As will quickly become evident, voices are extremely complex things, and some of the inherent limitations of the forensic-phonetic method are in part a consequence of the interaction between their complexity and the real world in which they are used. It is one of the aims of this book to explain how this comes about.

Because of the complexity of the subject matter, there is no straightforward way to present the information that is necessary to understand how voices can be related, or not, to their owners. I have chosen to organise this book into four main parts:

• The first part, in Chapters 2–5, introduces the basic ideas in FSI: Why voices are difficult to discriminate forensically, Forensic-phonetic parameters, Expressing the outcome, Characterising forensic-phonetic speaker identification.
• The second part, Chapters 6–9, describes what speech sounds are like: The human vocal tract and the production and perception of speech sounds, Phonemics, Speech acoustics, Speech perception.
• The third part, in Chapter 10, describes what a voice is.
• The fourth part, in Chapter 11, demonstrates the method using forensically realistic speech.

A more detailed breakdown of the contents of these four main parts – Basic ideas, What speech sounds are like, What is a voice? Forensic speaker identification demonstrated – is given below. But before that it is important to flag an additional feature of the overall organisation, namely that it reflects the strongly cumulative nature of the topic. The concepts in some early chapters constitute the building blocks for those to come. Thus, for example, it is difficult to understand how speakers differ in phonemic structure without understanding phonemics in Chapter 7, and very difficult to understand phonemics without a prior understanding of speech sounds from Chapter 6. The demonstration of the method in Chapter 11 will be much easier to understand with knowledge of speech acoustics in Chapter 8.

Because this book is intended to be of use to so many disparate groups, I have adopted an inclusive approach and erred on the side of detail. This means that some readers may find more detail in some chapters than they need. I can imagine that members of the legal profession will not be able to get quite as excited as I do about formant frequencies, for example, and readers with some background knowledge of articulatory phonetics do not need to learn about vocal tract structure. Feel free to skip accordingly, but please be aware of the cumulative structure of the book’s argument.

A brief characterisation of the contents now follows.

Basic ideas

In Chapters 2 to 5 are introduced the ideas that are central to the problem of FSI. There are four themes, one to each chapter. The first theme, in Chapter 2, describes what it is about voices that makes FSI difficult. This includes the existence of withinspeaker as well as between-speaker variation, and its consequences for discriminating between speakers; what conditions variation, and our lack of control over it in the real world. The second theme, in Chapter 3, is forensic-phonetic parameters. Here are discussed the different types of parameters used to compare speech samples forensically. The third theme describes the proper conceptual framework and way of expressing the outcome of a forensic-phonetic identification. This is in Chapter 4. The fourth theme, in Chapter 5, shows how FSI relates to other types of speaker recognition.

What speech sounds are like

FSI is performed on recordings of human vocalisations – sounds made exclusively by a human vocal tract. Although other vocalisations, for example laughs or screams, may from time to time be forensically important (Hirson 1995), most of the vocalisations used in FSI are examples of speech. Speech is the primary medium of that supremely human symbolic communication system called Language. One of the functions of a voice – perhaps the main one – is to realise Language, by conveying some of the speaker’s thoughts in linguistic form. Speech is Language made audible.

Moreover, when forensic phoneticians compare and describe voices, they usually do so with respect to linguistic units, especially speech sounds, like vowels or consonants. They might observe, for example, that the ee vowels in two samples are different; or that the th sound is idiosyncratically produced in both. It is therefore necessary to understand something of the structure of speech sounds, and how they are described. A large part of this book, Chapters 6 to 9, is accordingly given over to a description of the nature of speech sounds.

How speech sounds are produced – articulatory phonetics – is covered in Chapter 6. Since speech sounds are traditionally described in articulatory terms, their description is also covered in this chapter.

Chapter 7 (Phonemics) is concerned with how speech sounds are functionally organised in language. Phonemics is a conceptual framework that the forensic description of speech sounds usually assumes, and within which it is conveniently, and indeed indispensably, presented.

Phonemics regards actual speech sounds (called phones) as realisations of abstract sounds, called phonemes, whose function is to distinguish words. Thus the vowel in the word bead, and the vowel in the word bid are said to realise different phonemes in English because the phonetic difference between them – one vowel is, among other things, shorter than the other – signals a difference between words. The consonant sounds at the beginning of the words red and led realise different phonemes for the same reason. Phonemics is indispensable, because it supplies the basis for comparison of speech sounds in the first place. It allows us to say, for example, that two speech sounds are potentially comparable across samples because they are realisations of the same phoneme. This would enable us to legitimately compare, acoustically and phonetically, the vowel in the word car in one sample, say, with the vowel in the word far in another, because they are realisations of the same phoneme.

It is generally assumed that similarities and differences between forensic speech samples should be quantified acoustically, and acoustic comparison is an indispensable part of forensic-phonetic investigation. It is therefore necessary to describe speech acoustics, especially those that are assumed to be the ones in which speaker identity is optimally encoded. Speech acoustics are described in Chapter 8.

Several aspects of speech perception – how humans decode the speech acoustics to hear speech – are relevant for forensic phonetics. In particular, one argument for the necessity of analysing forensic speech samples both auditorily and acoustically is based on speech perception. It is also important to know about the expectation effect (you hear what you expect to hear), another aspect of speech perception. Both are discussed in Chapter 9: Speech perception.

The contents of Chapters 6 to 9, then, describe some of the basic knowledge that informs forensic-phonetic work. For example, it would be typical for a forensicphonetic expert to listen to two speech samples, decide what is comparable in terms of occurrences of the same phoneme, describe and transcribe the phonemes’ realisations, and then quantify the differences acoustically.

Each of these chapters on its own constitutes a vast area of scholarship. Scholarship, moreover, that is informed by several disciplines. Thus the multidisciplinary basis of phonetics comprises at least linguistics, anatomy, physiology, acoustics and statistics. Although I cannot hope to cover the contents of each chapter in a depth appropriate to the complexity of subject matter, I have tried to avoid the Charybdis of cursory treatment: to present here the figure of a spectrogram as an example of speech acoustics; there a word transcribed phonetically as an example of articulatory phonetics. Instead I have tried to include enough speech acoustics to help the reader appreciate how one speaker can differ acoustically from another, at least in some respects, and to understand what they are actually looking at in a spectrogram. I have also included enough traditional phonetics and phonology to help them understand at least some of what might otherwise be arcane parts of a forensicphonetic report. As a result of all this, the reader will find that they are able to follow important arguments demonstrated with the appropriate acoustics and linguistic analysis.

These aims have their downside. It would be silly to pretend that all of the specialist topics I have tried to explain in some detail are easy. They are quite definitely not, and the reader must be prepared to find at least some of the chapters difficult. Like many difficult passages, however, they become much more understandable on repeated readings.

What is a voice?

A voice is more than just a string of sounds. Voices are inherently complex. They signal a great deal of information in addition to the intended linguistic message: the speaker’s sex, for example, or their emotional state or state of health. Some of this information is clearly of potential forensic importance. However, the different types of information conveyed by a voice are not signalled in separate channels, but are convolved together with the linguistic message. Knowledge of how this occurs is necessary to interpret the ubiquitous variation in speech, and to assess the comparability of speech samples.

Familiar things like voices w...

Table of contents

• Cover Page
• Title Page
• Copyright Page
• Acknowledgements
• 1 Introduction
• 2 Why voices are difficult to discriminate forensically
• 3 Forensic-phonetic parameters
• 4 Expressing the outcome
• 5 Characterising forensic speaker identification
• 6 The human vocal tract and the production and description of speech sounds
• 7 Phonemics
• 8 Speech acoustics
• 9 Speech perception
• 10 What is a voice?
• 11 The likelihood ratio revisited: A demonstration of the method
• 12 Summary and envoi
• Glossary
• References