SincAlignNet

Implementation based on:
Frequency-Based Alignment of EEG and Audio Signals Using Contrastive Learning and SincNet for Auditory Attention Detection

SincAlignNet is an innovative framework for auditory attention detection that aligns EEG and audio features using an enhanced SincNet architecture with contrastive learning. It achieves state-of-the-art accuracy on KUL and DTU datasets, supporting efficient low-density EEG decoding for practical neuro-guided hearing aids.

---

Framework Overview

Fig. 1: SincAlignNet architecture for AAD, consisting of two phases:

1. Contrastive Learning - Aligns EEG and attended audio encodings by maximizing mutual information of correct EEG-Audio pairs
2. Inference - Identifies attended audio via cosine similarity between EEG/audio features or direct EEG-based inference

---

Encoder Architecture

Fig. 2: EEG and Audio encoder structure. Both encoders contain four components:

1. Multi-SincNet Bandpass
   • EEG: 60 filters | Audio: 320 filters
2. Depth Conv1D - Combines filter outputs for deeper features
3. Down Sample - Compresses data while preserving key information
4. Projector - Maps features to 128D latent space

---

Module Specifications

<img width="525" height="414" alt="image" src="https://github.com/user-attachments/assets/b74521b9-c58e-41f2-8865-4205f79812d5" />

Fig. 3: Component implementations:
(a) Depth-wise 1D convolution block
(b) Down sample module
(c) Projector architecture

---

Biological Motivation

Fig. 4: Proposed auditory attention mechanisms:

1. Noise Reduction (Fig 4a)

   • Brain processes mixed audio → extracts attended speaker
   • Simulated using SincNet filtering architecture

2. Information Minimization (Fig 4b)

   • Attentional focus minimizes mutual information entropy
   • Implemented via contrastive learning paradigm

---