Investigating the Phonetic Properties of Word Tokens in Interactional Contexts and the Implications for Forensic Voice Comparison

法医说话人比对（FSC）使用语音分析来协助刑事案件，通常涉及将罪犯的语音样本（例如辱骂语音邮件）与已知嫌疑人的样本进行比较，以评估两个样本由同一说话人制作的可能性。FSC依赖的特征在一个人的讲话中变化不大，但在说话者之间变化很大。FSC经常使用片段特征（辅音和元音）。对话分析（CA）领域发现人们如何相互交谈的模式。CA使用基于回合、功能和动作的定量数据。这两个领域都分析语音，但它们很少共享方法。话语和互动信息对FSC的潜在帮助研究很少。所进行的分析仅限于频率计数，例如不流畅。这项研究将更进一步，使用声学数据的跨学科方法。话语标记语是本研究理想的交互特征。dm是“空洞的表达”，经常出现，几乎没有命题意义，经常被污名化，简短，多功能，出现在许多位置（Brinton 2010）。例子包括“actually”和“I mean”，例如“我的意思是，实际上，我很累”。FSC通常涉及简短的语音样本，例如电话呼叫或录音错误，因此频繁的特性（如dm）是有用的。尽管频率很高，但DMs从未被分析用于识别说话人。然而，有相当多的证据表明，他们应该提供有用的特定于说话人的信息。对like的语音变化的研究显示出基于话语功能和周围语境的系统模式。Drager（2007）和Schleef & Turton（2016）分析了like作为动词、副词、引语(如：“她说‘没门’”)，DM和连词。他们比较了与话语位置的边界“强度”相关的声学测量。这些研究得出结论，动词标记比引语标记在元音中有更大的声学运动，但比引语标记少。因此，引号中的like比DM符号更接近“lark”。“like”周围更强的边界也会导致元音移动更少，/k/的读音也会减少。填充停顿（FPs，即呃，嗯）的声学分析也产生了说话者特定的模式（Hughes et al. 2016）。本文将使用CA方法将“like”和FPs的分析扩展到各种dm和其他单词；以我的硕士论文为基础分析将参考每个标记的功能，旋转位置和旋转功能。采用更全面、跨学科的方法，将揭示更多关于dm和演讲者使用它们的方式，同时也为FSC案例工作的方法提供信息。这些考虑导致了一系列的研究问题：RQ 1: dm的发音在哪些方面有所不同？rq2：说话者是否以特定于说话者的方式一致地发出dm，从而使它们成为有用的法医特征？rq3：改变上下文（即功能和转动位置）如何影响说话者的DMs发音？rq4：这些上下文是否增加了DMs的法医潜力？为了回答rq1 -3，我将研究白玫瑰大学举办的一系列语音语料库中dm的语音变化。分析将涉及定量语音方法，包括统计模型，如最近的通用加性混合模型，我在我的硕士论文中成功地将其应用于“like”和“yeah”。我还将从CA收集定性见解，这将告知如何解释每个单词令牌。这项研究将是一个创新的尝试，将CA和FSS结合起来，以扩大法医方法和社会语音学研究的范围。它将直接告知如何利用以前未被探索的语言方面来帮助法医案件工作，从而使言论作为证据的司法问题更加清晰。

Forensic speaker comparison (FSC) uses phonetic analysis to aid criminal cases and typically involves comparing a speech sample made by a criminal (e.g. abusive voicemails) with a sample from a known suspect to assess the likelihood of the two samples being made by the same speaker. FSC relies on features which have low variation within an individual's speech but are highly variable between speakers. FSC often uses segmental features (consonants & vowels). The field of Conversation Analysis (CA) finds patterns in how people talk to each other. CA uses quantitative data based on turns, function and action.These fields both analyse speech, but they rarely share methodologies. The potential of discourse and interactional information to aid FSC has been little researched. What analysis has been conducted is limited to frequency counts, for example of disfluencies. This study will go further, using interdisciplinary methods on acoustic data. Discourse markers (DMs) are ideal interactional features for this research. DMs are "empty expressions" which appear frequently, have little propositional meaning and are often stigmatized, short, multifunctional and seen in many positions (Brinton 2010). Examples include actually and I mean as in "I mean, actually, I'm very tired". FSC often involves short speech samples, e.g. phone calls or recording bugs, therefore frequent features such as DMs are useful.Despite their frequency, DMs have never been analysed for the purposes of speaker identification. However, there is considerable evidence that they should provide useful speaker-specific information. Research into the phonetic variation of like shows systematic patterning based on utterance function and surrounding context. Drager (2007) and Schleef & Turton (2016) analysed the differences between like as a verb, adverb, quotative (e.g. "she was like 'no way'"), DM and conjunction. They compared acoustic measures related to boundary 'strength' of the discourse position. These studies concluded that verb tokens have greater acoustic movement within the vowel than quotative tokens, but less than DMs. Thus, quotative occurrences of like sound closer to 'lark' than DM tokens. A stronger boundary surrounding 'like' also leads to less vowel movement and a reduction of /k/. Acoustic analysis of filled pauses (FPs; i.e. uh, um) has also yielded speaker-specific patterns (Hughes et al. 2016).This thesis will extend the analysis of 'like' and FPs to a variety of DMs and other words using CA methods; building on my MSc dissertation. Analysis will refer to each token's function, turn position and the turn function. Taking a more holistic, interdisciplinary approach will unveil more about DMs and the way speakers use them whilst also informing the methods of FSC casework.These considerations lead to a series of research questions: RQ 1: In what ways do the pronunciations of DMs vary? RQ 2: Do speakers pronounce DMs consistently, and in a speaker-specific way, thereby making them useful forensic features? RQ 3: How does changing the context (i.e. function and turn position) impact the way speakers pronounce DMs? RQ 4: Do these contexts increase the forensic potential of DMs? To answer RQs 1-3, I will study the phonetic variation of DMs in a range of speech corpora held at White Rose universities. The analysis will involve quantitative phonetic methods, including statistical models such as the recent General Additive Mixed Models, which I successfully applied to 'like and 'yeah' in my MSc dissertation. I will also gather qualitative insights from CA which will inform how each word token will be interpreted.This research will be an innovative attempt at uniting CA and FSS to broaden the scope of forensic methods and the sociophonetic study of DMs. It will directly inform how a previously unexplored aspect of speech can be used to aid forensic casework, thereby adding clarity to issues of justice wherever speech is a piece of evidence.