ÄÄNEEN PERUSTUVA KÄYTTÄJÄN TODENTAMINEN PUHELINVERKON LISÄARVOPALVELUISSA

Tässä diplomityössä perehdytään puhujantunnistukseen ja sen käyttökelpoisuuteen käyttäjän henkilöllisyyden todentamisessa osana puhelinverkon lisäarvopalveluja. Puhelimitse ohjattavat palvelut ovat yleensä perustuneet puhelimen näppäimillä lähetettäviin äänitaajuusvalintoihin. Käyttäjän henkilöllisyydestä on voitu varmistua esimerkiksi käyttäjätunnuksen ja salaisen tunnusluvun perusteella. Tulevaisuudessa palvelut voivat perustua puheentunnistukseen, jolloin myös käyttäjän todentaminen äänen perusteella vaikuttaa järkevältä. Työssä esitellään aluksi erilaisia biometrisiä tunnistamismenetelmiä. Työssä perehdytään tarkemmin äänen perusteella tapahtuvaan puhujan todentamiseen. Työn käytännön osuudessa toteutettiin puhelinverkon palveluihin soveltuva puhujantodennussovelluksen prototyyppi. Työn tarkoituksena oli selvittää teknologian käyttömahdollisuuksia sekä kerätä kokemusta puhujantodennuspalvelun toteuttamisesta tulevaisuutta silmällä pitäen. Prototyypin toteutuksessa ohjelmointikielenä käytettiin Javaa.

Speaker diarization: Segmentation and clustering of speeches

Speaker diarization is the process of sorting speeches according to the speaker. Diarization helps to search and retrieve what a certain speaker uttered in a meeting. Applications of diarization systemsextend to other domains than meetings, for example, lectures, telephone, television, and radio. Besides, diarization enhances the performance of several speech technologies such as speaker recognition, automatic transcription, and speaker tracking. Methodologies previously used in developing diarization systems are discussed. Prior results and techniques are studied and compared. Methods such as Hidden Markov Models and Gaussian Mixture Models that are used in speaker recognition and other speech technologies are also used in speaker diarization. The objective of this thesis is to develop a speaker diarization system in meeting domain. Experimental part of this work indicates that zero-crossing rate can be used effectively in breaking down the audio stream into segments, and adaptive Gaussian Models fit adequately short audio segments. Results show that 35 Gaussian Models and one second as average length of each segment are optimum values to build a diarization system for the tested data. Uniting the segments which are uttered by same speaker is done in a bottom-up clustering by a newapproach of categorizing the mixture weights.

New kernel functions and learning methods for text and data mining

Recent advances in machine learning methods enable increasingly the automatic construction of various types of computer assisted methods that have been difficult or laborious to program by human experts. The tasks for which this kind of tools are needed arise in many areas, here especially in the fields of bioinformatics and natural language processing. The machine learning methods may not work satisfactorily if they are not appropriately tailored to the task in question. However, their learning performance can often be improved by taking advantage of deeper insight of the application domain or the learning problem at hand. This thesis considers developing kernel-based learning algorithms incorporating this kind of prior knowledge of the task in question in an advantageous way. Moreover, computationally efficient algorithms for training the learning machines for specific tasks are presented. In the context of kernel-based learning methods, the incorporation of prior knowledge is often done by designing appropriate kernel functions. Another well-known way is to develop cost functions that fit to the task under consideration. For disambiguation tasks in natural language, we develop kernel functions that take account of the positional information and the mutual similarities of words. It is shown that the use of this information significantly improves the disambiguation performance of the learning machine. Further, we design a new cost function that is better suitable for the task of information retrieval and for more general ranking problems than the cost functions designed for regression and classification. We also consider other applications of the kernel-based learning algorithms such as text categorization, and pattern recognition in differential display. We develop computationally efficient algorithms for training the considered learning machines with the proposed kernel functions. We also design a fast cross-validation algorithm for regularized least-squares type of learning algorithm. Further, an efficient version of the regularized least-squares algorithm that can be used together with the new cost function for preference learning and ranking tasks is proposed. In summary, we demonstrate that the incorporation of prior knowledge is possible and beneficial, and novel advanced kernels and cost functions can be used in algorithms efficiently.

Equivalence and translation strategies in multilingual thesaurus construction

In the present dissertation, multilingual thesauri were approached as cultural products and the focus was twofold: On the empirical level the focus was placed on the translatability of certain British-English social science indexing terms into the Finnish language and culture at a concept, a term and an indexing term level. On the theoretical level the focus was placed on the aim of translation and on the concept of equivalence. In accordance with modern communicative and dynamic translation theories the interest was on the human dimension. The study is qualitative. In this study, equivalence was understood in a similar way to how dynamic, functional equivalence is commonly understood in translation studies. Translating was seen as a decision-making process, where a translator often has different kinds of possibilities to choose in order to fulfil the function of the translation. Accordingly, and as a starting point for the construction of the empirical part, the function of the source text was considered to be the same or similar to the function of the target text, that is, a functional thesaurus both in source and target context. Further, the study approached the challenges of multilingual thesaurus construction from the perspectives of semantics and pragmatics. In semantic analysis the focus was on what the words conventionally mean and in pragmatics on the ‘invisible’ meaning - or how we recognise what is meant even when it is not actually said (or written). Languages and ideas expressed by languages are created mainly in accordance with expressional needs of the surrounding culture and thesauri were considered to reflect several subcultures and consequently the discourses which represent them. The research material consisted of different kinds of potential discourses: dictionaries, database records, and thesauri, Finnish versus British social science researches, Finnish versus British indexers, simulated indexing tasks with five articles and Finnish versus British thesaurus constructors. In practice, the professional background of the two last mentioned groups was rather similar. It became even more clear that all the material types had their own characteristics, although naturally not entirely separate from each other. It is further noteworthy that the different types and origins of research material were not used to represent true comparison pairs, and that the aim of triangulation of methods and material was to gain a holistic view. The general research questions were: 1. Can differences be found between Finnish and British discourses regarding family roles as thesaurus terms, and if so, what kinds of differences and which are the implications for multilingual thesaurus construction? 2. What is the pragmatic indexing term equivalence? The first question studied how the same topic (family roles) was represented in different contexts and by different users, and further focused on how the possible differences were handled in multilingual thesaurus construction. The second question was based on findings of the previous one, and answered to the final question as to what kinds of factors should be considered when defining translation equivalence in multilingual thesaurus construction. The study used multiple cases and several data collection and analysis methods aiming at theoretical replication and complementarity. The empirical material and analysis consisted of focused interviews (with Finnish and British social scientists, thesaurus constructors and indexers), simulated indexing tasks with Finnish and British indexers, semantic component analysis of dictionary definitions and translations, coword analysis and datasets retrieved in databases, and discourse analysis of thesauri. As a terminological starting point a topic and case family roles was selected. The results were clear: 1) It was possible to identify different discourses. There also existed subdiscourses. For example within the group of social scientists the orientation to qualitative versus quantitative research had an impact on the way they reacted to the studied words and discourses, and indexers placed more emphasis on the information seekers whereas thesaurus constructors approached the construction problems from a more material based solution. The differences between the different specialist groups i.e. the social scientists, the indexers and the thesaurus constructors were often greater than between the different geo-cultural groups i.e. Finnish versus British. The differences occurred as a result of different translation aims, diverging expectations for multilingual thesauri and variety of practices. For multilingual thesaurus construction this means severe challenges. The clearly ambiguous concept of multilingual thesaurus as well as different construction and translation strategies should be considered more precisely in order to shed light on focus and equivalence types, which are clearly not self-evident. The research also revealed the close connection between the aims of multilingual thesauri and the pragmatic indexing term equivalence. 2) The pragmatic indexing term equivalence is very much context-depended. Although thesaurus term equivalence is defined and standardised in the field of library and information science (LIS), it is not understood in one established way and the current LIS tools are inadequate to provide enough analytical tools for both constructing and studying different kinds of multilingual thesauri as well as their indexing term equivalence. The tools provided in translation science were more practical and theoretical, and especially the division of different meanings of a word provided a useful tool in analysing the pragmatic equivalence, which often differs from the ideal model represented in thesaurus construction literature. The study thus showed that the variety of different discourses should be acknowledged, there is a need for operationalisation of new types of multilingual thesauri, and the factors influencing pragmatic indexing term equivalence should be discussed more precisely than is traditionally done.

Power Consumption Measurement and Scenario Based Energy Model in Wearable Computing Applications

The power is still today an issue in wearable computing applications. The aim of the present paper is to raise awareness of the power consumption of wearable computing devices in specific scenarios to be able in the future to design energy efficient wireless sensors for context recognition in wearable computing applications. The approach is based on a hardware study. The objective of this paper is to analyze and compare the total power consumption of three representative wearable computing devices in realistic scenarios such as Display, Speaker, Camera and microphone, Transfer by Wi-Fi, Monitoring outdoor physical activity and Pedometer. A scenario based energy model is also developed. The Samsung Galaxy Nexus I9250 smartphone, the Vuzix M100 Smart Glasses and the SimValley Smartwatch AW-420.RX are the three devices representative of their form factors. The power consumption is measured using PowerTutor, an android energy profiler application with logging option and using unknown parameters so it is adjusted with the USB meter. The result shows that the screen size is the main parameter influencing the power consumption. The power consumption for an identical scenario varies depending on the wearable devices meaning that others components, parameters or processes might impact on the power consumption and further study is needed to explain these variations. This paper also shows that different inputs (touchscreen is more efficient than buttons controls) and outputs (speaker sensor is more efficient than display sensor) impact the energy consumption in different way. This paper gives recommendations to reduce the energy consumption in healthcare wearable computing application using the energy model.