Fundamental Frequency Distributions of Bilingual Speakers in Forensic Speaker Comparison (2017)
Type of ContentTheses / Dissertations
Degree NameMaster of Linguistics
PublisherUniversity of Canterbury
AuthorsDorreen, Kieranshow all
The fundamental frequency (f0) is the main acoustic correlate of the voicing pitch produced by a speaker. Since the beginning of forensic speaker comparison research f0 has been a popular measure, however its highly variable nature has led to many years of debate on its usefulness as a speaker discriminant. Nolan (1983) suggests six parameters to be used in speaker comparison: 1. High between-speaker variability 2. Low within-speaker variability 3. Resistance to attempted disguise or mimicry 4. Availability 5. Robustness in transmission 6. Measurability
Nolan argues that f0 is an accurate speaker discriminant, asserting that it complies with the above requirements. He lauds especially its availability, measurability, and its robustness. However, there is a large debate on f0’s effectiveness when it comes to the first two parameters: between-speaker variability and within-speaker variability. F0 may be a common, easily measurable discriminant, but it is also shown to be highly variable within-speaker. This alone would make it useless as a speaker discriminant as two different voice samples of the same speaker could potentially have two very different f0 measurements. Despite this f0 is still one of the most popular speaker discriminants used by forensic phoneticians (Gold 2014).
Forensic phoneticians have until recently largely ignored what may be a substantial portion of the fundamental frequency range (see section 2.3). Creak phonation sits in a speaker’s lower frequency range, a range that is often too low for pitch tracking tools to accurately discover. However, with new advancements in the accuracy of pitch tracking it may now be possible to accurately measure creak within a speaker. This has major implications for the analysis of f0 as an entire phonation type that was essentially ignored and / or mismeasured can now be systematically evaluated.
In this thesis I will present a new way to quantify the fundamental frequency within forensic speaker comparison. In chapter 2 I will outline the current state of fundamental frequency analysis in the field of forensic speaker comparison. F0 analysis has been marred by inconsistent practices and inaccurate data collection methods stemming from the use of imprecise pitch tracking. To rectify these methodological inconsistencies I will propose a new method of f0 analysis in chapter 3. This new method starts with a new, more robust pitch tracker that can accurately track the lower frequency ranges. This new method see creak phonation data being analyzed alongside the modal f0 data in a way that was impossible before without accurate pitch tracking. I will test this new method in chapter 4, looking at the within-speaker variation and the between-speaker variation of two sets of bilingual speakers: a homogenous group of bilingual speakers and a heterogeneous group of bilingual speakers. The recordings of these bilingual speakers are ideal for examining f0 variation both within-speaker and betweenspeaker as all conditions except the language they speak in remain the same. Chapter 5 will present small case studies of three speakers who had substantially different frequency distributions than the rest of the speakers in the two corpuses. I will discuss the implications these outliers may have on f0 analysis and forensic speaker recognition. And finally in chapter 6 I will summarize the results of the proposed method of f0 analysis and suggest a way forward for f0 analysis in forensic speaker comparison.