The human voice is a very efficient communication medium that we depend on each day to transfer messages. By taking a closer look, it turns out that the speech we produce is never just a message, but always contains additional, more unconsciously perceived cues about the speaker, the so-called paralinguistic information. The field of Speaker Classification, a sub-field of artificial intelligence, computational linguistics, signal processing, and others, is concerned with the extraction of this information from the speech signal. Examples of such speaker characteristics that manifest themselves in the voice are the person’s age, gender, physics, emotion, personality, attention, cognitive load, or even identity. The main goal of this dissertation is to develop a novel Speaker Classification system for automatic age and gender recognition that allows this information be used in applications for various purposes. One particular motivation is drawn from the aim to utilize the nonintrusively obtained knowledge for personalization of functions and services in the automotive context, which is a very active domain for human-machine interaction research. One aspect of this thesis deals with the question of how current Speaker Classification systems can be improved. The methodology consequently follows a principled approach, motivating the investigation through analyses of how the speech apparatus is affected by age and gender, instead of looking only at the decision problem. An extensive study and comparison of several recent and early approaches points out where research is heading, what the most promising leads are, and where the culprits are located. Based on this survey, an approach named Frisc is developed and presented in detail. It combines the promising GMM-SVM supervector pattern recognition system with frame-based cepstral voice features (MFCCs) and applies further optimizations to source, feature, and score domain. To be a suitable choice for time-critical on-line application, the system was further designed to process incoming data and update knowledge in incremental manner and with real-time efficiency. A further focus of this principled approach lies in the way the evaluation of the system is performed. As part of the contribution of this work, a carefully designed parameter space exploration experiment is conducted to discover the impact of each of a large number of different parameters to the overall classification performance. Parameters include MFCC extraction settings, GMM training parameters, SVM kernel function, background model impact, feature and silence filtering ratio, and more. Best practices with respect to data set generation and balancing are particularly emphasized. In spite of the complexity introduced, the procedure ensures that the complete system is always tested using a large number of samples, and only a single parameter is changed between any two experiments. It is believed that the very specific results presented in this thesis are more valuable in the long run than an unstructured or less

A speaker classification framework for non-intrusive user modeling : speech-based personalization of in-car services

Cyclists are among the most vulnerable road users. Many recent interventions have aimed at improving their safety on the road, such as the minimum overtaking distance rule introduced in Queensland in 2014. Smartphones offer excellent opportunities for technical intervention for road safety at a limited cost. Indeed, they have a lot of available processing power and many embedded sensors that allow analysing a rider's (or driver's) motion, behaviour, and environment; this is especially relevant for cyclists, as they do not have the space or power allowance that can be found in most motor vehicles. The aim of the study presented in this paper is to assess cyclists’ support for a range of new smartphone-based safety technologies. The preliminary results for an online survey with cyclists recruited from Bicycle Queensland and Triathlon Queensland, with N=191, are presented. A number of innovative safety systems such as automatic logging of incidents without injuries, reporting of dangerous area via a website/app, automatic notification of emergency services in case of crash or fall, and advanced navigation apps were assessed. A significant part of the survey is dedicated to GoSafeCycle, a cooperative collision prevention app based on motion tracking and Wi-Fi communications developed at CARRS-Q. Results show a marked preference toward automatic detection and notification of emergencies (62-70% positive assessment) and GoSafeCycle (61.7% positive assessment), as well as reporting apps (59.1% positive assessment). Such findings are important in the context of current promotion of active transports and highlight the need for further development of system supported by the general public.

Using smartphones for cycling safety: A survey of riders preferences and interest in new technologies

A prototypical conspicuity enhancement (CE) system for vulnerable road users (here e-bikes) is described. We stress that CE is a form of multimodal output. We argue that previous CE approaches have the drawback of affecting uninvolved (road) users. We argue further that augmented reality as an alternative is error prone because objects need to be tracked. Our system implements the hybrid reality modality model, where directed information emanates from the objects themselves and therefore no object recognition/tracking is needed. We describe the components of a functional demonstrator based on standard compliant car-to-car communication components.

Nils Schnabel

Papers

Multimodal conspicuity-enhancement for e-bikes: the hybrid reality model of environment transformation