Quantitative approaches to problems in linguistics

This volume is a collection of papers by friends and colleagues of Phil Rose to honour him for his significant contribution to the field of linguistics on the occasion of his recent retirement from the ANU. Phil has influenced us all in Illany important and lasting ways. His work in both (especially Chinese) tone and forensic voice comparison has always nmllagcd to improve upon the accepted standard through the insistence all statisti cal signifi cance and the introduction of statistical techniques not previously employed in the field . He is a true sc ientist and always makes sure that his work was both ambitious and of the highest calibre. Phil has a long history of working on tona l phenomena, since his Master's thesis on the phonology of Ningbo (Chinese). Despite having somc of the best cars around, Phil quickly identified the shortcomings of auditory impressions for seriolls tona l descriptions, turning to the aeollstie signal for data that can bc propcrly quantified and measured (e.g. Rose 1982). He used many speakers for his early studies to ensure taking the mcan eliminatcd performance quirks. However, Phi l had also secn efforts to normalize acoustic data for vowels and had the great idea of extending the genera l practise to tona l descripti ons as we ll. Needless to say, Phil's 1987 paper, "Considerations on the normalisation of the fundamental frequency of linguistic tone", had a major impact on the field . Phil used a simple statistic - the z-scorc transfoml - to normalize his acousti c data of the tones to abstract awuy from the variation found between speakers, making il possibl e 10 identify the lanai contou rs representative of the variety as (I whole. He has sincc made studies of many varieties of Chinese (and other languages), clarifying isslies for tonological rcscareh as well as providing quality quantificd descriptions for the fi eld of linguistics. Phi l's approach is now the gold standard for tona l descriptions, and 110 doubt p!lvcd the way for his work in forensic voice comparison. Phi l's work on fo rensic voice compmison (FYC) has turned the field Oil its head, quite literally. In a lypic,)1 FVC easc, you havc a speech recording from an unknown voice (e.g. an offender), and a speech sample from a known subject (c.g. a suspect). Beforc the introduction of the likelihood ratio-based approach, the problem of FYC was considered a classification/identification task; dctcrmining, for example, whether the specch from the offender and suspect we re simi lar enough to be identificd as coming from the same person (or more "precisely", the probabi lity of this being the case). In tcchnica l terms, this is equiva lent to trying to eva lu ate the probabili ty of a hypothesis (c.g. the sllspect is guilty), given Ihe evidencc (c.g. the observed similarities and differenccs between speech pallems of the suspect and the offender). Indeed, FYC was long considered a subfield of automatic speaker recognition and even referred to as forensic speaker recognition/ identificmion/verification. However, Phil has shown that there arc serious problems (both technical and Icga l) with the o ld approach. Not only is it impossiblc for the forcnsic specialist to directly calculate the (conditiona l) probabilities of such hypotheses (e.g. the suspec t is guilty) given the evidence available to him/her, but also, statemen ts on the probabi lity of such hypotheses being true potentially change the scientist's ro le from that of cxpert witness, to onc of " trier-of-fac t", norma lly reserved for judges and juries.