I study the speech and language of individuals with Primary Progressive Aphasia and Alzheimer’s Disease to identify objective measures that can enable their diagnosis early, provide prognosis of their condition, and ultimately improve therapy. To this end, I employ natural language processing, acoustic analysis, and machine learning models (deep neural networks, SVMs, etc.).
|Themistocleous Charalambos, Bronte Ficek, Kimberly Webster, Dirk-Bart den Ouden, Argye E. Hillis, Kyrana Tsapkini (2021). Automatic subtyping of individuals with Primary Progressive Aphasia. Journal of Alzheimer’s Disease, https://doi.org/10.3233/JAD-201101 In this paper, we present a machine learning model based on deep neural networks (DNN) for the subtyping of patients with PPA into three main variants, using combined acoustic and linguistic information elicited automatically. The end-to-end automated machine learning approach we present can enable clinicians and researchers to provide an easy, quick, and inexpensive classification of patients with PPA.|
|Themistocleous Charalambos (2019). Dialect Classification from a Single Sonorant Sound Using Deep Neural Networks. doi: 10.3389/fneur.2018.00975. Listeners do not require long productions of speech to identify the accent of a speaker, often a single sound suffices. This study shows that using machine learning and information from a speech segment, namely a single sonorant sound /m n l r/, it is possible to distinguish two dialects of Greek: Athenian Greek and Cypriot Greek. In our future research, we will be exploring further this approach to identify medical conditions that influence speech production.|
|Themistocleous Charalambos, Eckerström Marie, and Dimitrios Kokkinakis (2018). Identification of Mild Cognitive Impairment from Speech in Swedish using Deep Sequential Neural Networks. Frontiers in Neurology. doi: 10.3389/fneur.2018.00975.To this day, there is no cure for dementia but early-stage treatment can delay the progression of MCI; thus, the development of valid tools for identifying early cognitive changes is of great importance. In this study, we provide an automated machine learning method, using Deep Neural Network Architectures, that aims to identify MCI. The Deep Neural Network Architecture proposed here constitutes a method that contributes to the early diagnosis of cognitive decline, quantifies the progression of the condition, and enables suitable therapeutics.
Frontiers in Neurology 2018 version: [PDF]
Link to GitHub page with source code: [CODE]
|Themistocleous Charalambos (2017). Dialect classification using vowel acoustic parameters. Speech Communication 94, 13 -22. This study provides a classification model of two Modern Greek dialects, namely Athenian Greek and Cypriot Greek, using information from speech. To this end, a large corpus of vowels from 45 speakers of Athenian Greek and Cypriot Greek was collected. The findings show 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.Link to the paper: [PDF]|
|Themistocleous Charalambos (2017). The Nature of Phonetic Gradience across a Dialect Continuum: Evidence from Modern Greek Vowels. Phonetica 74, 157–172. This study investigates the acoustic properties of vowels in Athenian Greek (AG) and Cypriot Greek. The findings show that (1) stressed vowels are more peripheral than unstressed vowels, (2) AG unstressed /i a u/ vowels are more raised than the corresponding CG vowels, (3) AG unstressed vowels are shorter than CG unstressed vowels, and (4) AG /i·u/ are more rounded than the corresponding CG vowels.Link to the paper: [PDF]|
|Themistocleous Charalambos (2017). Effects of two linguistically proximal varieties on the spectral and coarticulatory properties of fricatives: Evidence from Athenian Greek and Cypriot Greek. Frontiers in Psychology. The central thesis of this paper is that cross-dialectal studies of fricative's acoustic structure can reveal patterns that designate speakers of different dialectal groups. The findings provide a solid evidence base for the manifestation of dialectal information in the acoustic structure of fricatives. Frontiers 2017 Version: [PDF]|
|Grohmann Kleanthes, Papadopoulou Elena and Themistocleous Charalambos (2017). Acquiring Clitic Placement in Bilectal Settings: Interactions between Social Factors. Frontiers in Communication. The C5.0 machine-learning algorithm was employed to model the interaction of sociolinguistic factors on the development of clitic placement in bidialectal children. The model shows that speakers acquire the relevant features very early, yet compartmentalization of form and function according to style emerges only as they engage in the larger speech community. Frontiers 2017 Version: [PDF]|
|Themistocleous Charalambos (2016). The bursts of stops can convey dialectal information. Journal of the Acoustical Society of America EL 140(4), EL334–EL340.This study investigates the effects of the dialect of the speaker on the spectral properties of stop bursts. Forty-five female speakers—20 Standard Modern Greek and 25 Cypriot Greek speak- ers—participated in this study. The spectral properties of stop bursts were calculated from the burst spectra and analyzed using spectral moments. The findings show that besides linguistic information, i.e., the place of articulation and the stress, the speech signals of bursts can encode social information, i.e., the dialects. A classification model using decision trees showed that skewness and standard deviation have a major contribution for the classification of bursts across dialects. Link to the JASA paper: [PDF]|
|Themistocleous Charalambos (2016). Seeking an anchorage: Evidence from the tonal alignment of the Cypriot Greek prenuclear pitch accent. Language and Speech, 59(4):433–461.By exploring the timing of the Cypriot Greek L*+H prenuclear pitch accent, this study tested the predictions of three hypotheses about tonal alignment: the invariance hypothesis, the segmental anchoring hypothesis, and the segmental anchorage hypothesis. The findings on the alignment of the high tone (H) are both intriguing and unexpected: the alignment of the H depends on the number of unstressed syllables that follow the prenuclear pitch accent. The `wandering' of the H over multiple syllables is extremely rare among languages, and casts doubt on the invariance hypothesis and the segmental anchoring hypothesis, as well as indicating the need for a modified version of the segmental anchorage hypothesis. To address the alignment of the H, we suggest that it aligns within a segmental anchorage–the area that follows the prenuclear pitch accent–in such a way as to protect the paradigmatic contrast between the L*+H prenuclear pitch accent and the L+H* nuclear pitch accent. Link to the Language and Speech paper: [PDF]|
|Themistocleous Charalambos (2014). Edge-Tone Effects and Prosodic Domain Effects on Final Lengthening. Linguistic Variation 14(1). 129–160This study reports two experiments that investigate the edge-tones and domain-specific effects on final lengthening. The study shows that in Cypriot Greek the following occur: (a) lengthening applies primarily on the syllable nucleus not the syllable onset, which suggests variety specific effects of lengthening; (b) lengthening depends on the edge-tones, namely, polar questions trigger more lengthening than statements and wh-questions; (c) lengthening provides support for at least two distinct prosodic domains over the phonological word, the intonational phrase and the intermediate phrase; greater lengthening associates with the first and shorter lengthening with the latter; (d) finally, syllable duration depends on the syllable distance from the boundary. By pointing to the distinct lengthening effects of edge-tones and domain-boundaries, the aforementioned findings highlight the application of different lengthening devices.Link to the PDF paper: [PDF]|