Body-Brain Waves - 28th September '24 - Talk by Bernd Accou

Body-Brain Waves - 28th September '24 - Talk by Bernd Accou

by Bernd Accou

Applied Sciences & Engineering Neuroscience PSALife sciences: general issuesPSANNeurosciences
🎙 Bernd Accou 👥 23 📅 December 5, 2024 ⏱ 14 min 👁 14 📄 expert opinion 🧭 2026-08-18
Available in: English (current) Français

Keywords

speech intelligibilitydeep learningEEGhearing testslistening engagement

Summary

Bernd Accou, from the ExpORL lab, presents his PhD research on using deep learning models to predict speech intelligibility and listening engagement from EEG. He begins by highlighting the global burden of hearing loss and the limitations of current hearing tests, which rely on behavioral responses and unnatural stimuli. His approach uses natural speech (audiobooks) and EEG to train a deep learning model to reconstruct the speech envelope, correlating it with the actual envelope as a measure of intelligibility. He compares this physiological measure to the standard Matrix test, showing a significant correlation in estimating the speech reception threshold. He acknowledges that his model, trained on a separate group, performs slightly worse than a subject-specific linear model but offers practical advantages. Future work includes integrating multiple bodily signals (ECG, GSR, pupil diameter) and using multimodal deep learning and multitask pre-training to improve listening engagement prediction. He also promotes the publicly available PIZ dataset for training such models. The talk concludes with a Q&A session discussing model architecture, explainability, and potential confounds like eye movements.

174 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the application of deep learning to a real-world clinical problem, offering a promising alternative to traditional hearing tests. The argumentation is solid, grounded in the limitations of current methods and the potential of deep learning to address them. The speaker clearly explains the methodology and presents results comparing his model to a standard test, demonstrating its validity. However, the talk is more of an overview of ongoing research rather than a detailed presentation of a single study, and some claims could benefit from more rigorous statistical details.

102 words

Title / Content Match

The title accurately reflects the talk's focus on brain-body interactions and the specific topic of predicting speech intelligibility.

Quality & Reliability

7/10

The talk presents ongoing research with a clear methodology and comparison to a standard test, but lacks detailed peer-reviewed references and is a conference presentation.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • PIZ dataset — Mentioned as a publicly available dataset for training deep learning models on EEG and speech.

Concurring Sources

  • PIZ dataset — The dataset is publicly available and supports the research presented.

Contribution & Novelties

The talk presents a novel application of deep learning to predict speech intelligibility from EEG, potentially enabling objective hearing tests without behavioral responses. The approach uses natural stimuli and cross-subject generalization, which are significant improvements over traditional methods. The future integration of multiple physiological signals could lead to more robust measures of listening engagement.

Pour aller plus loin :

  • Deep learning for EEG-based speech decoding — Relevant to the methodology and dataset.
  • Speech envelope reconstruction from EEG — Foundational work in the field.
  • Multimodal deep learning — Relevant to the proposed integration of multiple signals.

95 words

Radar Profile

The radar profile shows balanced scores across all dimensions, indicating a well-rounded presentation with solid information, technical depth, and reliability, though not exceptional in any single area.

Reliability 7/10