- Note: I created a GitHub repository with selected papers, which can be accessed here.
Deep Neural Networks, SVMs, HMMs constitute the powerhouse of most applications in text-to-speech, speech perception, machine translation, image description generation, and semantic interpretation. They also offer insights about human perception and cognition, namely how humans process, store, retrieve information from speech signals, texts, vision, etc. I explore these architectures to understand how humans map speech into abstract linguistic categories.
Specifically, in Themistocleous (2017), I provide a classification model of two Modern Greek dialects, namely Athenian Greek and Cypriot Greek, using information from formant dynamics of F1, F2, F3, F4 and vowel duration. The measurements were employed in classification experiments, using three classifiers: Linear Discriminant Analysis, Flexible Discriminant Analysis, and C5.0. The latter outperformed the other classification models, resulting in a higher classification accuracy of the dialect. C5.0 classification shows that duration and the zeroth coefficient of F2, F3 and F4 contribute more to the classification of the dialect than the other measurements; it also shows that formant dynamics are important for the classification of dialect.
The ability to accurately perceive whether a speaker is asking a question or is making a statement is crucial for any successful interaction. However, learning and classifying tonal patterns has been a challenging task for automatic speech recognition and for models of tonal representation, as tonal contours are characterized by significant variation. In Bernardi and Themistocleous (2017) we provide a classification model of Cypriot Greek questions and statements. We evaluate two state-of-the-art network architectures: a Long Short-Term Memory (LSTM) network and a convolutional network (ConvNet). The ConvNet outperforms the LSTM in the classification task and exhibited an excellent performance with 95% classification accuracy.