Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting gamma-band responses to the speech envelope for the ICASSP 2024 Auditory EEG Decoding Signal Processing Grand Challenge

Mike Thornton, Jonas Auernheimer, Constantin Jehn, Danilo Mandic, Tobias Reichenbach

TL;DR

This work tackles the match-mismatch EEG decoding problem in speech perception by extending prior approaches to use gamma-band responses to the speech envelope. It introduces a dual-branch architecture with a low-frequency envelope-tracking decoder and a gamma-band envelope decoder, both trained with randomized mismatched segments and sharing a common deep-learning backbone. The two logits are fused with linear discriminant analysis in a composite decoder, which yields the best performance. On the ICASSP SPGC dataset, gamma-band evidence of envelope tracking is demonstrated, and ensemble averaging further boosts accuracy, highlighting the practical potential for robust multi-band auditory EEG decoding.

Abstract

The 2024 ICASSP Auditory EEG Signal Processing Grand Challenge concerns the decoding of electroencephalography (EEG) measurements taken from participants who listened to speech material. This work details our solution to the match-mismatch sub-task: given a short temporal segment of EEG recordings and several candidate speech segments, the task is to classify which of the speech segments was time-aligned with the EEG signals. We show that high-frequency gamma-band responses to the speech envelope can be detected with a high accuracy. By jointly assessing gamma-band responses and low-frequency envelope tracking, we develop a match-mismatch decoder which placed first in this task.

Detecting gamma-band responses to the speech envelope for the ICASSP 2024 Auditory EEG Decoding Signal Processing Grand Challenge

TL;DR

This work tackles the match-mismatch EEG decoding problem in speech perception by extending prior approaches to use gamma-band responses to the speech envelope. It introduces a dual-branch architecture with a low-frequency envelope-tracking decoder and a gamma-band envelope decoder, both trained with randomized mismatched segments and sharing a common deep-learning backbone. The two logits are fused with linear discriminant analysis in a composite decoder, which yields the best performance. On the ICASSP SPGC dataset, gamma-band evidence of envelope tracking is demonstrated, and ensemble averaging further boosts accuracy, highlighting the practical potential for robust multi-band auditory EEG decoding.

Abstract

The 2024 ICASSP Auditory EEG Signal Processing Grand Challenge concerns the decoding of electroencephalography (EEG) measurements taken from participants who listened to speech material. This work details our solution to the match-mismatch sub-task: given a short temporal segment of EEG recordings and several candidate speech segments, the task is to classify which of the speech segments was time-aligned with the EEG signals. We show that high-frequency gamma-band responses to the speech envelope can be detected with a high accuracy. By jointly assessing gamma-band responses and low-frequency envelope tracking, we develop a match-mismatch decoder which placed first in this task.
Paper Structure (4 sections, 2 figures, 2 tables)

This paper contains 4 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. Results
  4. Conclusions

Figures (2)

  • Figure 1: High-gamma ([range-phrase=--]70220Hz, red) and lower-gamma ([range-phrase=--]35150Hz, blue) envelope-related EEG temporal response functions. The high-gamma range captures the typical range of the fundamental frequency of speech. The global field potentials (GFPs) shown in the figure have been normalised so that the background activity has a value of 1. Therefore, the peak GFP values represent the SNRs of the two responses.
  • Figure 2: Comparison between the output logits of the low-frequency (LF) and gamma-band envelope decoders, as well as the decision boundary of the composite decoder. Orange dots represent matched segments, whereas red dots represent mismatched segments.