Description of Work
T2.1 – Investigating existing practices (M1-M5)
The purpose of this task is threefold as it aims to the investigation of the most suitable methods for imagined speech decoding, the understanding of the underlying neural mechanisms of imagined speech and the identification of publicly available EEG-based datasets for imagined speech decoding, that are considered as crucial aspects towards the development of informed decoding algorithms. Extra attention will be paid to Deep Learning approaches given the nature of the task, but also considering that such approaches will be more efficient, robust and easy to adapt to the needs of the incremental vocabulary generation (see T2.3).
T2.2 – Neuro-informed imagined speech decoding algorithms (M6-M21)
This task aims at the design of neuro-informed algorithms for imagined speech decoding. The developed algorithms will leverage both the knowledge gained as part of T2.1 and the team’s expertise in the decoding process of similar mental tasks (e.g. motor imagery, mental arithmetic). The neuro-informed algorithms will be firstly evaluated on existing datasets (identified by T2.1) and then adapted for the in-house dataset generated as part of T3.2.
T2.3 – Incremental Learning for Vocabulary Expansion (M8-M21)
T2.3 aims at enriching existing vocabularies with extra prompts. Within this context, incremental learning approaches will be employed to investigate the potential of multiclass scalability of the algorithmic frameworks developed in T2.2, with the main target being the robust extensibility of the imagined speech vocabulary.
T2.4 – Interconnection between languages (M15-M21)
This task will explore the semantics of the imagined speech, aiming to identify commonalities and differences between words of identical meaning in different languages (i.e., Greek and English). This may shed light to the existing neuroscientific hypotheses with respect to the imagined speech paradigm or even give rise to new ones. Finally, we must note here that this task is highly dependent on tasks T3.1 and T3.2 as to the best of our knowledge publicly available datasets that contain bilingual mental imagery speech data do not exist.
T2.5 – Imagined Speech decoding Toolbox (M15-M24)
This task will gather the different methods, modules and pre-trained models developed and employed as part of WP2 (i.e. T2.2 –T2.4) in an open-source, properly documented and ready-to-use toolbox, accessed via a publicly available repository.