Project

Discipline

Phonetics; Prosody; Speaker Identification; Forensics; Timing; Rhythm

Lead
Lay summary
Everyday experiences tell us that it is typically possible to identify a speaker solely on the basis of his/her voice (e.g. when someone starts a phone call with a simple 'hi' or when people talk in a different room). Such observations reveal that speakers carry individual features in their voices by which they can be identified to a considerable degree. The present project aims at studying the role of temporal characteristics of the speech signal in speaker identification. The study will pay particular attention to possible applications of the results in the field of forensic phonetics in which phonetic knowledge is applied in legal cases where the identity of the speaker in a recording is disputed. We start from the observation that the acoustic speech signal is made up of dynamic processes resulting from the movements of the articulators. It has been demonstrated successfully in other scientific domains that humans can be identified on the basis of their movements only, e.g. by the way they walk. Our working hypothesis is that the movements of the organs of speech (e.g. jaw, lips or tongue) can be equally idiosyncratic as human gait and that idiosyncratic ways to move the organs of speech leave individual temporal charcateristics in the acoustic speech signal. We will therefore study numerous durational parameters in speech from segment durations (e.g. the durations of consonants and vowels) over syllable and word to prosodic durations (e.g. durational characteristics of intonation). In the first year of the project we are aiming at identifying temporal measures of speech that are most speaker-idiosyncratic. In the second year we will test these measures towards within speaker variability (e.g. different types of voice disguise). In year three we will use behavioral experimental methods to test whether the measures we have identified as being most speaker-idiosyncratic are perceptually salient (i.e. whether listeners can identify a speaker solely on the basis of certain temporal voice characteristics). It is well possible that we will find that some temporal speaker idiosyncratic features are perceptually salient and others are not. We argue that the salient temporal features will help us explaining how human listeners identify speakers on the basis of their voice. Non-salient features, however, may be less prone to within speaker variability like voice disguise as they should be difficult to control for speakers. Such features may thus be of high value for acoustic voice identification of non-cooperative speakers (i.e. speakers not wishing to be identified) typically found under forensic circumstances.

Direct link to Lay Summary

Last update: 21.02.2013

Active participation

Active participation

Communication	Title	Media	Place	Year

Number	Title	Start	Funding scheme

Speakers' voices are to a high degree individual but we only have a limited understanding of this phenomenon. The major theoretical aim of the present project is to investigate how temporal characteristics of human speech (e.g. segmental or prosodic timing patterns, speech rhythmic characteristics and durational patterns of voicing) contribute to speaker individuality. On a practical level we will investigate how knowledge about temporal differences between speakers can be applied to forensic phonetic speaker identification. Speaker identification, in particular in forensic phonetic contexts, has predominantly been carried out on the basis of spectral characteristics of a speaker’s voice (e.g. fundamental frequency of vocal fold vibration and vocal tract resonances like vocalic formant frequencies). It has been argued correctly that such frequency content is directly influenced by idiosyncratic anatomical features of a speaker’s organs of speech (in particular the size of the larynx and lengths of the vocal tract cavities) which limit the range of certain spectral parameters and can thus contribute to making speakers’ voices individual. The emphasis is on ‘contribute’ as experience has taught us that there are clear limits in identifying speakers based on spectral parameters alone. It is therefore necessary to explore other dimensions in speech where idiosyncratic information is encoded. Such a dimension is ‘time’ and it has been paid surprisingly little attention to in the past. This is surprising because research from other domains, such as motion pattern recognition, has demonstrated convincingly that humans have highly individual ways in which they move and that individuals can be identified, for example, by means of temporal gait information alone. In the present project we argue that speech is similar to walking in that it is a highly complex brain operated control mechanism over a large number of muscle movements which may all be carried out to some degree in individual ways. We then go one step further and argue that such idiosyncratic motion does not need to be observed visually from articulator movement itself but can be found in the acoustic speech signal, as this is the immediate product of all speech articulatory movements. To study temporal individuality in speech we will therefore (a) systematically analyze the durational characteristics that vary most across speakers of the Standard German variety spoken in Zürich and explain the reasons for temporal variability between speakers. We will (b) test how robust such characteristics are towards sources of within-speaker variability (e.g. voice disguise or varying emotional content of speech) and between speaker similarity (e.g. speakers imitating each other). We will finally (c) test whether between-speaker temporal differences are perceptually salient. We argue that non-salient speaker idiosyncratic temporal characteristics are most valuable for acoustic forensic speaker identification as speakers should have limited control over manipulating these parameters for the case they are trying to hide their identity (i.e. in voice disguise).