Deliverables

BINGO is deeply engaged in the collection and analysis of EEG data to decode imagined speech, contributing to advancements in Brain-Computer Interfaces (BCIs). To achieve this efficiently, the project implements structured data acquisition and evaluation protocols.The present document outlines the data generation and management processes involved in BINGO’s experimental studies and analyses. It provides a detailed description of how EEGs are collected, at what stages, for what specific purposes, along with the methods employed for data processing. Additionally, the report defines the protocols for data flow, backup procedures, and secure storage, ensuring compliance with ethical and legal frameworks. Given the sensitive nature of neural data, the document also establishes measures to safeguard participant privacy and maintain strict data protection standards. Furthermore, an assessment of BINGO’s alignment with FAIR (Findability, Accessibility, Interoperability, and Reusability) principles is conducted to enhance data transparency and reusability. As a living document, the Data Management Plan (DMP) will be continuously reviewed and updated throughout the project’s duration to reflect technological advancements, regulatory changes, and evolving scientific needs.

Deliverable D2.1 ‘Report on the review of imagined speech decoding approaches’ summarizes the existing practices with respect to EEG-based imaged speech decoding with the aim to identify the most suitable methods and publicly available EEG-based datasets for imagined speech decoding necessary for the development of informed decoding algorithms. The report outlines the challenges of Brain Computer Interface (BCI) systems, the advancements in EEG signal processing and the shift towards deep learning techniques in data processing; specifically, deep learning techniques such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to assess temporal and spatial dependencies within EEG signals.

Decoding imagined speech from electroencephalography (EEG) signals represents a significant advancement in brain-computer interface (BCI) research, offering new possibilities for assistive communication and neurorehabilitation. This report presents an in-depth investigation of three distinct decoding approaches: (i) Riemannian Geometry-based analysis for feature extraction and denoising, (ii) a hybrid deep learning framework combining EEGNet with Riemannian Geometry, and (iii) an attention-based EEG Conformer model. These methods are evaluated using publicly available datasets, incorporating neuroscientific principles from the dual-stream model of speech processing to improve decoding accuracy. The findings indicate that Riemannian Geometry methods enhance the discrimination of imagined speech patterns by leveraging spatial covariance matrices, while the EEGNet-Riemannian hybrid improves classification performance through end-to-end feature learning. The EEG Conformer, despite its promise in capturing long-range dependencies, exhibits challenges related to generalization and overfitting. Across all approaches, subject variability and the difficulty of distinguishing phonetically similar words remain significant hurdles. The code for the methodologies described in this report are available at the project’s code repository

This deliverable investigates cross-linguistic neural phenomena associated with inner speech using electroencephalography (EEG), with the aim of identifying language-invariant and language-dependent brain representations. Two complementary methodological approaches are explored: a Riemannian geometry framework that models EEG activity through covariance-based representations on symmetric positive-definite manifolds, and a spectro-temporal analysis that examines frequency- and time-resolved neural dynamics. Due to the absence of suitable publicly available EEG datasets for systematic cross-linguistic inner speech research, the proposed methodologies were applied to the BINGO cross-linguistic inner speech dataset, which was specifically collected for this purpose. The analyses presented provide insight into multilingual inner speech processing while establishing a methodological foundation for future cross-language EEG studies.

This document accompanies BINGO’s GitHub and presents a detailed overview of the Imagined Speech Decoding Toolbox. The repository comprises a comprehensive toolbox that supports EEG-based imagined speech research across the entire processing pipeline, from experimental data acquisition to preprocessing, segmentation, and downstream data analysis.

This document constitutes the first version of the D3.1, which describes the project’s recording protocol. In detail it includes: i) the objectives that drive overall design process, ii) the experimental setup including both the hardware and software that will realize the protocol, iii) the recording protocol by delineating the prompts, the stimuli, and the experiment’s timeline, and iv) the information regarding the participants with emphasis on the inclusion and exclusion criteria.

This document accompanies the recorded data that have been gathered as part of BINGO and constitutes a sequel of D3.1-v1. The collected data aim to cover two basic aspects of inner speech decoding research: i) A complete and large-scale dataset that supports a wide variety of words (also capable of acting as a generic spelling corpus), and ii) A dataset capable of driving research among the neural interconnections between inner speech in two distinct languages. In this document we present the final experimental protocols, and the appropriate quality metrics that guarantee the validity of the recorded data. This deliverable aims to serve as a document accompanying the released datasets (available at project’s website).
The

This deliverable investigates EEG-based imagined speech decoding within the BINGO project, aiming to assess decoding performance under varying levels of task complexity and generalization. EEG data were collected from 20 participants over three sessions, using a 26-word vocabulary based on the NATO phonetic alphabet. Multiple evaluation protocols were employed, including within-session and cross-session subject-dependent experiments, as well as subject-independent evaluations using leave-one-subject-out and Monte Carlo cross-validation.
Two deep learning architectures were examined: EEGNet, serving as a compact baseline model, and an EEG Conformer, which integrates convolutional layers with self-attention mechanisms. Results show that decoding performance is highest in within-session settings, decreases in cross-session evaluations, and approaches chance level under subject-independent conditions. Reduced-vocabulary experiments yield more stable performance for both models. Overall, the EEG Conformer consistently matches or slightly outperforms EEGNet, particularly in more challenging evaluation scenarios, while absolute performance highlights the persistent difficulty of imagined speech decoding from EEG signals. The code for the methodologies described in this report are available at the project’s code repository.

The D4.1 – Project Communication Kit, outlines the communication and dissemination strategy for the BINGO project. BINGO aims to develop a Brain-Computer Interface (BCI) system capable of decoding imagined speech through advanced machine learning techniques and a neuro-informed approach. To maximize impact and engagement, this deliverable presents the key communication tools designed for the project: a project leaflet, the official website, and social media channels. These tools will facilitate effective outreach, ensuring visibility among researchers, industry stakeholders, and the broader public. The document details the structure, content, and objectives of each communication component, demonstrating how BINGO will enhance knowledge dissemination and collaboration within the scientific and technological communities.

The Dissemination and Communication Plan (D4.2) outlines the strategies, methods, and  activities designed to maximize the impact of the project through effective communication and  continuous dissemination efforts. The plan details how BINGO will engage with key target audiences, including the scientific community, policymakers, industry professionals, and the general public, ensuring that research outcomes are effectively communicated to all relevant stakeholders. Central to the plan is the identification of stakeholders and the development of tailored communication channels, tools, and dissemination activities, all of which will be executed and monitored throughout the project’s duration. All research team members will be actively involved in these communication efforts, with responsibilities shared across the related tasks. Additionally, BINGO will seek collaboration and exchange with related projects to amplify its impact and explore synergies, ensuring that the results of the project reach their full potential. The plan also emphasizes the importance of feedback mechanisms and the monitoring of dissemination activities to measure the effectiveness of the communication strategy. Furthermore, the project aims to demonstrate the value of EU funding by fostering networks of stakeholders, establishing team members hips with industry, and contributing to the broader research community through conferences, workshops, and other events. Ultimately, this Communication and Dissemination Plan ensures that the project’s outcomes will be widely shared, understood, and used to drive future research, collaboration, and exploitation beyond the project’s completion.

This deliverable reports on the scientific competition and impact assessment of the BINGO project at Month 24 (M24). It provides a structured and evidence-based overview of communication, dissemination, and uptake activities implemented during the project, alongside documentation of the design and submission of a scientific competition as a complementary dissemination instrument. The assessment covers scientific and societal dimensions of impact, drawing on documented dissemination outputs, conference participation, digital communication activities, and monitored engagement indicators. Particular emphasis is placed on the organisation and submission of the BINGO NATO Alphabet dataset to the Kaggle platform (decision still pending), which has the potential to establish an open and reproducible benchmarking framework for electroencephalography (EEG)-based imagined speech decoding. The deliverable demonstrates that, by M24, BINGO has implemented a coherent, monitored, and adaptive approach to communication, dissemination, and scientific competition activities, resulting in verifiable scientific outputs, sustained outreach, and the establishment of infrastructure supporting ongoing scientific engagement.