https://www.um.edu.mt/library/oar/handle/123456789/11952 2026-06-25T09:24:08Z 2026-06-25T09:24:08Z https://www.um.edu.mt/library/oar/handle/123456789/78357 2021-07-15T10:37:31Z 2016-01-01T00:00:00Z

Title: A coarticulation model for articulatory speech synthesis Abstract: The state-of-the-art techniques for speech synthesis rely either on concatenation of acoustic units taken from a vast pre-recorded speech database noting the relevant linguistic information or on statistical generation of the necessary acoustic parameters and using a speech production model. These approaches yield synthesis of good quality, but are purely technical solutions which bring no or very little information about the acoustics of speech or about how the articulators (mandible, tongue, lips, velum...) are controlled. In contrast, the articulatory approach generates the speech signal from the vocal tract shape and its modelled acoustic phenomena. The vocal tract deformation control comprises slow anticipation of the main constriction and fast and imperatively accurate aiming for consonants. The system predicts the sequence of vocal tract consecutive configurations from a sequence of phonemes of the French language to be articulated and a model of the coarticulation effects in it. We use static magnetic resonance imaging (MRI) captures of the vocal tract shape when producing phonemes in various contexts, thus following an approach by Birkholz (2013). The evaluation of the model is done both on the animated graphics representing the vocal tract shape evolution (how natural and efficient the movement is) and on the synthesised speech signals that are perceptively and-in terms of formants-qualitatively compared to identical utterances made Ly a human. Our results show that there are a lot of effects in the dynamic process of speech that manage to be reproduced by manipulating solely static data. We discuss generation of pure vowels, vowel-to-vowel and vowel-consonant-vowel transitions, and articulators' behaviour in phrases, report which acoustic properties have been rendered correctly and what could be the reasons for the system to fail to produce the desired result in other cases, and ponder how to reduce the after-effects of target-oriented moves to obtain a more gesture-like motion. Description: M.SC.ARTIFICIAL INTELLIGENCE

2016-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/72835 2024-10-28T10:49:10Z 2016-01-01T00:00:00Z

Title: Behaviour mining for personalised desktop tool-support Abstract: It is natural for individuals to think about their work in terms of tasks, where a task could be the documents that they are working on or with, and/or the directories containing documents related to the same task. But directories do not necessarily contain all of the related documents, because documents might be emails and web pages, which are harder to place into the same directory. Furthermore, individuals are multitaskers and tasks get interrupted. This results in task-switching and the eventual resumption of some task (possibly the same one). Although different applications provide tools to support individuals to manage their information space, these tools fall short in supporting individuals when they need to keep and re-find documents that are part of their task or when a task-switch occurs. Task-keeping and re-finding (and/or resuming) with the current tool-support philosophy is both time-consuming and limited since it does not reflect the task concept. In contrast to a myriad of other approaches, this thesis considers task-keeping as a process that keeps track of changing tasks by identifying which documents belong to which tasks and when a task-switch occurs. An attempt is made to automatically identify the documents that belong to a task by solely relying on the user’s switching and re-visitation behaviour. In this way the process aims to reduce the keeping time without introducing new interruptions. Prior to addressing the problem, an experiment is conducted in a controlled environment with 22 participants to collect ground-truth task-related data. This dataset is later used to evaluate the task-keeping solution. It is extensively analysed from different perspectives using a visual-analytics tool that we built (called PiMx), to draw up the design requirements for the incremental density-based graph-clustering algorithm iDeTaCt. The algorithm is evaluated and extended and a marked improvement in its performance registered. A usability study is also performed with 15 participants over a five-day period using a prototype task-re-finding and resumption tool called PiMxT that uses iDeTaCt to automatically keep tasks. Both qualitative and quantitative feedback are solicited and analysed to verify the usefulness of the approach. The results show that the time to re-find and resume tasks with PiMxT is reduced by almost a half in the majority of the cases considered and that there is considerable potential behind this combined approach to task-keeping, re-finding and resumption. Description: PH.D.ARTIFICIAL INTELLIGENCE

2016-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/12238 2024-04-18T11:36:21Z 2016-01-01T00:00:00Z

Title: News event time-lines Abstract: As the amount of news portals increases, huge amounts of news stories are being published every hour which leads to an information overload problem, and readers cannot digest it all. When reading a news report about an event, readers would find it useful to be able to conceptualise how that event unfolded without needing to read through all the previous news reports. This research focuses on implementing a live system News Event Time-lines(NETs) that creates a visualization interface that uses automatically generated time-lines from news sources to represent news events. The process involved many different stages, starting from the extraction of news stories, feature extraction, clustering into news events, and the timeline visualizations. The purpose of this system is to show the evolution of news events and how they relate with each other. A user study was conducted where participants made use of NETs showed promising results about the usefulness of the system, if it were to be extended further. Description: B.SC.IT(HONS)

2016-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/12183 2018-04-16T14:20:18Z 2016-01-01T00:00:00Z

Title: Automatic song genre classification using visual features from the spectrogram Abstract: Due to the advent of the world wide web and hypermedia support, data is being generated and consumed by users at significantly high rates. Among the many types of data is audio data – music in particular – and music files can come with various pieces of information associated with them, such, as name, author, album, and genre. These are called ID3 tags, and for a long time, the pursuit towards automation of the creation of these ID3 tags has been under way. The goal of this project is to focus on one of the aforementioned ID3 tags: the genre. Simply put, a genre is a label used to classify songs under pre-determined categories, and the system will aspire to automatically classify a given song under a genre. The area of genre classification has largely been focused on using the audio domain to carry out the analysis and feature vectors required. While this may seem at first like the logical approach, most research in the area has hit a glass ceiling where improvement on existing results is difficult. Clearly, more research into a newer approach would revitalise the area and work towards improving on such results. Hence, the approach that this paper will take makes use of image processing, or more precisely, visually analysing the spectrogram (time vs frequency graph), which is a more unique approach towards genre classification. Description: B.SC.IT(HONS)

2016-01-01T00:00:00Z