In the last decade, research in Speech Emotion Recognition (SER) has become a major endeavour in Human Computer Interaction (HCI), and speech processing. Accurate SER is essential for many applications, like assessing customer satisfaction with quality of services, and detecting/assessing emotional state of children in care. The large number of studies published on SER reflects the demand for its use. The main concern of this thesis is the investigation of SER from a pattern recognition and machine learning points of view. In particular, we aim to identify appropriate mathematical models of SER and examine the process of designing automatic emotion recognition schemes. There are major challenges to automatic SER including ambiguity about the list/definition of emotions, the lack of agreement on a manageable set of uncorrelated speech-based emotion relevant features, and the difficulty of collected emotion-related datasets under natural circumstances. We shall initiate our work by dealing with the identification of appropriate sets of emotion related features/attributes extractible from speech signals as considered from psychological and computational points of views. We shall investigate the use of pattern-recognition approaches to remove redundancies and achieve compactification of digital representation of the extracted data with minimal loss of information. The thesis will include the design of new or complement existing SER schemes and conduct large sets of experiments to empirically test their performances on different databases, identify advantages, and shortcomings of using speech alone for emotion recognition. Existing SER studies seem to deal with the ambiguity/dis-agreement on a “limited” number of emotion-related features by expanding the list from the same speech signal source/sites and apply various feature selection procedures as a mean of reducing redundancies. Attempts are made to discover more relevant features to emotion from speech. One of our investigations focuses on proposing a newly sets of features for SER, extracted from Linear Predictive (LP)-residual speech. We shall demonstrate the usefulness of the proposed relatively small set of features by testing the performance of an SER scheme that is based on fusing our set of features with the existing set of thousands of features using common machine learning schemes of Support Vector Machine (SVM) and Artificial Neural Network (ANN). The challenge of growing dimensionality of SER feature space and its impact on increased model complexity is another major focus of our research project. By studying the pros and cons of the commonly used feature selection approaches, we argued in favour of meta-feature selection and developed various methods in this direction, not only to reduce dimension, but also to adapt and de-correlate emotional feature spaces for improved SER model recognition accuracy. We used rincipal Component Analysis (PCA) and proposed Data Independent PCA (DIPCA) by training on independent emotional and non-emotional datasets. The DIPCA projections, especially when extracted from speech data coloured with different emotions or from Neutral speech data, had comparable capability to the PCA in terms of SER performance. Another adopted approach in this thesis for dimension reduction is the Random Projection (RP) matrices, independent of training data. We have shown that some versions of RP with SVM classifier can offer an adaptation space for Speaker Independent SER that avoid over-fitting and hence improves recognition accuracy. Using PCA trained on a set of data, while testing on emotional data features, has significant implication for machine learning in general. The thesis other major contribution focuses on the classification aspects of SER. We investigate the drawbacks of the well-known SVM classifier when applied to a preprocessed data by PCA and RP. We shall demonstrate the advantages of using the Linear Discriminant Classifier (LDC) instead especially for PCA de-correlated metafeatures. We initiated a variety of LDC-based ensembles classification, to test performance of scheme using a new form of bagging different subsets of metafeature subsets extracted by PCA with encouraging results. The experiments conducted were applied on two benchmark datasets (Emo-Berlin and FAU-Aibo), and an in-house dataset in the Kurdish language. Recognition accuracy achieved by are significantly higher than the state of art results on all datasets. The results, however, revealed a difficult challenge in the form of persisting wide gap in accuracy over different datasets, which cannot be explained entirely by the differences between the natures of the datasets. We conducted various pilot studies that were based on various visualizations of the confusion matrices for the “difficult” databases to build multi-level SER schemes. These studies provide initial evidences to the presence of more than one “emotion” in the same portion of speech. A possible solution may be through presenting recognition accuracy in a score-based measurement like the spider chart. Such an approach may also reveal the presence of Doddington zoo phenomena in SER.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:681209 |
| Date | January 2015 |
| Creators | Al-Talabani, Abdulbasit |
| Publisher | University of Buckingham |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://bear.buckingham.ac.uk/101/ |
Page generated in 0.0024 seconds