Boxing Day, Leeds

• On 26 December 2003, two traffic officers stopped a car with false number plates. The driver shot at them and escaped, and PC Ian Broadhurst was killed
• A camera on the police car’s dashboard recorded the gunman’s voice, but not his face

The known samples

• The suspect, David Bieber, was an American who had lived in Yorkshire for some years under a false identity
• After his arrest he said very little in police interviews, so the police used other known samples: his calls to telephone betting companies

A mixed accent

Both recordings had the same unusual mixture:

• Rhotic /r/, pronounced in all positions (right, pardon; road, third): typical of North American accents, not of Yorkshire
• Typical Yorkshire FACE and GOAT vowels, and Yorkshire terms of address such as mate and love: not found in American accents

Why it was clear-cut

• The recordings were similar: the same features in both
• The combination was atypical: very few people in the relevant population speak like that
• And the key features were easy to hear: no acoustic measurement was needed

Today

1. Why comparing voices is hard
2. Which features are ideal, and which features experts analyse
3. Automatic speaker recognition
4. How experts express conclusions: likelihood ratios

Forensic speaker comparison

• A questioned recording (an unknown speaker) is compared with a known recording (a suspect, usually from a police interview)
• Questioned recordings: ransom demands, drug deals arranged by phone, a stalker’s voice messages
• Speaker comparison makes up about 70% of forensic speech casework
• In the UK, it was first used in evidence in 1965, and now in around 500 cases a year

Five approaches

1. Auditory analysis: careful, repeated listening
2. Acoustic analysis: measuring the signal with software
3. Auditory and acoustic analysis combined
4. Automatic speaker recognition (ASR)
5. ASR with human analysis

There is no “voiceprint”

• 1960s–1980s: the idea that spectrograms could be matched by eye, like fingerprints
• The FBI used it for years. But in a survey of 2,000 FBI comparisons, examiners reached a decision in only about a third (Koenig, 1986)
• Speech scientists rejected it: voices vary too much, and recordings differ too much

No two samples are identical

• Like a signature, speech is a dynamic movement: no one makes exactly the same movements twice
• Forensic recordings add more variation:
  • speaker: emotion, intoxication, a formal police interview, a taciturn suspect
  • technical: background noise, telephone transmission, reverberation, very short samples
• All of this can change within a single recording

So: no “match”

• Two recordings will always show some differences and some similarities
• The expert assesses how strongly the evidence supports one explanation over the other, and reports this on a scale

What makes a feature ideal?

• High between-speaker variation: speakers differ from each other
• Low within-speaker variation: stable across the day, health, situations
• Robust to transmission: survives the telephone
• Robust to disguise and mimicry
• And practical: frequent, easy to measure

Similarity and typicality

• Similarity: how alike are the questioned and known samples?
• Typicality: how common is the feature in the relevant population?
• The relevant population usually comes from a profile of the questioned speaker: for example, adult men from New Zealand
• Sometimes it is fixed by the case: if CCTV shows three people in the car, the population is the two others

Similarity and typicality

Simulated population of 500 men (mean 4.4 syllables per second). Q: questioned sample, S: suspect sample.

Reading the four cases

1. Neither similar nor typical, at opposite ends: points towards different speakers
2. Similar but typical: weak support for the same speaker, since many people speak this fast
3. Similar and atypical: relatively strong support for the same speaker
4. Not very similar, both within the typical range: no strong evidence either way

The telephone effect

Simulated. Telephones cut energy below about 350 Hz, which raises measured F1, most in low-F1 vowels: up to about 14% on landlines (Künzel, 2001), 29% on average on mobiles (Byrne & Foulkes, 2004).

Many layers

Speech is componential: it has layers that are largely independent of each other.

• Segments: vowels, consonants, connected speech processes
• Suprasegmentals: voice quality, f0, intonation, tempo, rhythm
• Higher-level features: words, grammar, discourse markers, pauses
• Non-linguistic features: breathing, laughing, throat clearing

Vowels

• Almost everyone analyses vowels, because formants depend on both anatomy and accent
• They are usually well-defined, even in poor recordings
• Formants can be measured at the midpoint, or at several points to capture formant dynamics: speakers move between targets in individual ways (McDougall, 2006)

Formant dynamics

Simulated F1 of five speakers in a close-to-open diphthong, like Thai /ia/. The speakers aim at the same vowels but move between them differently: E starts low and rises late, C rises early.

Vowels: caveats

• Telephone transmission raises F1, and louder speech and shouting raise formants in general
• So values will not simply “match”: the analyst interprets them, allowing for known sources of difference

Consonants

Often analysed by ear. In British English cases, for example:

• /h/-dropping: head as [ɛd]
• /θ ð/ as [f v] or as stops
• /t/ between vowels: [t], [ɾ] or [ʔ] in better
• -ing as [ɪŋ] or [ɪn]
• /l/ as [l], [ɫ] or a vowel; /r/ present or absent, and its form

Connected speech processes

Shortcuts in fluent speech, used differently by accent and by speaker:

• Assimilation: handbag as [ˈhamˌbag]
• Elision: library as [ˈlaɪbɹi]
• Epenthesis: hamster as [ˈhampstə]

Voice quality

• Most experts consider it, often with Vocal Profile Analysis (Laver, 1980), usually in simplified versions (San Segundo et al., 2019)
• Vocal tract settings (lips, jaw, tongue, larynx height), muscular tension, phonation type
• Needs substantial training, and ideally at least 40 seconds of speech
• Phonation is affected by stress, exhaustion, alcohol and emotion: a difference does not necessarily mean different speakers

Fundamental frequency

• Mean, range and standard deviation; at least about one minute of speech for representative values
• Robust to noise, and survives the telephone: f0 can be reconstructed from the harmonics. Harmonics at 500, 600, 700 and 800 Hz imply an f0 of 100 Hz
• But f0 varies a lot within a speaker: louder on the phone, higher in read speech, lower in the morning, higher when agitated
• Most useful for eliminating suspects, and easy to explain to a court

f0 in a population

Simulated. In the real Swiss German data (500 men, spontaneous speech), the median is 126 Hz and the range 81–197 Hz (Hedegard et al., 2023). Data like these let us judge typicality.

Tempo and rhythm

• Articulation rate varies less within a speaker than speaking rate
• But research data often come from students. In UK police interviews, suspects sometimes read a statement prepared by their lawyers: often slower and less fluent than their spontaneous speech
• Rhythm: the percentage of vocalic intervals (%V) is speaker-specific and robust to speaking style (Leemann, Kolly & Dellwo, 2014)

Higher-level features

• Pausing is speaker-specific and fairly stable
• Filled pauses (uh, um): how often, and their acoustics. Their acoustics discriminate speakers well, especially um (Hughes, Wood & Foulkes, 2016)
• Tag questions (innit?, right?), discourse markers (like), ways of greeting or answering the phone
• Probably below the level of consciousness, so hard to disguise

How ASR works

1. Pre-processing: resampling, voice activity detection
2. Feature extraction: usually mel frequency cepstral coefficients (MFCCs)
3. Speaker modelling: today, x-vectors from deep neural networks
4. Comparison of the questioned and known models: a score
5. Score interpretation: converting the score into a likelihood ratio, against a model of the relevant population

MFCCs

• Not phonetic features like f0 or formants, but compact descriptions of the whole spectrum, peaks and troughs
• The mel scale warps frequency like human hearing: finer at low frequencies
• Taken from short windows (20–30 ms), 10–20 coefficients every 10 ms or so: thousands per recording
• The resulting speaker model is sometimes called a “voiceprint”, with a different meaning from the 1960s term

Strengths

• More objective, though humans still choose the recordings, settings and population
• Replicable: same material and settings, same result
• Fast: seconds instead of hours
• In a 2019 survey, over 40% of practitioners used ASR (Gold & French, 2019)

Weaknesses

• Needs at least about 10 seconds of net speech per sample; 30 seconds is better
• Channel mismatch: a telephone call against a recording in a reverberant interview room. Newer x-vector systems cope better
• A black box: results cannot be explained in terms a court understands, such as pitch or accent
• When the book was written, ASR had not yet been admitted in evidence in UK courts

Discussion

An ASR system gives strong support for the same speaker. Your phonetic analysis finds that the questioned speaker has a northern English accent and the suspect a southern one.

How should the report deal with the disagreement? Would your answer change if the court had never heard ASR evidence before?

From “match” to strength of evidence

• Before the 1990s, conclusions were often binary (the same speaker, or not). This is now widely rejected, because no speech feature is unique or invariant
• Most practitioners now express conclusions as likelihood ratios (LRs), as recommended for forensic science generally (AFSP, 2009)

Crossing the road

• Evidence: the traffic you see, its speed, the width of the road
• Two hypotheses: I can cross safely; I cannot
• You weigh how likely the evidence is under each, and cross if the first clearly outweighs the second
• Other information also matters: how much of a hurry you are in, your experience of this road

The likelihood ratio

LR = p(evidence | same speaker) / p(evidence | different speakers)

• How likely is the evidence if the samples come from the same speaker, relative to different speakers?
• LR above 1: support for the same-speaker hypothesis. Below 1: for the different-speaker hypothesis
• An LR of 100: the evidence is 100 times more likely under one hypothesis than the other

A shoe print

• A size 13 print at a crime scene. The suspect wears size 13: similarity = 1
• About 5% of men wear size 13: typicality = 0.05
• LR = 1 / 0.05 = 20: the print is 20 times more likely if the suspect left it than if a random man did

Computing an LR

Hypothetical data. The questioned value (dashed): 0.491 / 0.0044 = LR of 111, log₁₀LR = 2.05.

Which population?

Hypothetical data. The same suspect and the same questioned value (dashed), against two definitions of the relevant population. A value that is unusual among all adult men can be common among men with the same accent, and the LR shrinks. Who belongs in the population depends on the facts of the case.

Why logarithms?

• An LR of 100 is as strong as an LR of 0.01, in the other direction
• Large numbers and tiny fractions are hard to compare
• log₁₀ makes the scale symmetrical around 0: 1,000,000 becomes 6, 0.000001 becomes −6
• Our example: LR 111 → log₁₀LR 2.05 → moderately strong support for the same speaker

A real case

• Australia, 2012–2013: fraud calls to a financial institution were recorded (Enzinger, Morrison & Ochoa, 2016)
• Relevant population: adult male speakers of Australian English, from a corpus
• The questioned call had noise, reverberation and landline transmission. The experts degraded the corpus recordings to match
• Result: LR = 3,287 (log₁₀LR = 3.52): strong support for the same-speaker hypothesis

Discussion

Is an LR of 3,287 big or small?

• What else would you need to know before deciding whether the suspect made the calls?
• Who should make that decision?

Exercises

For the practical sessions (book activities 4A and 4B):

• Measure f0 and articulation rate in one questioned and five suspect speakers. Which suspect is most similar, and is the similarity typical?
• Compare a telephone call with a police interview recording, and write a short report on at least four features

Summary

Slides and code

These slides

mcanzi.de/ws27_fss/04_speaker_comparison.html

Code and all sessions

github.com/mcanzi/ws27_fss