CA-TCN: A Causal-Anticausal Temporal Convolutional Network for Direct Auditory Attention Decoding

Iñigo García-Ugarte; Rubén Eguinoa; Ricardo San Martín; Daniel Paternain; Carmen Vidaurre

cs.SDMar 27, 2026

CA-TCN: A Causal-Anticausal Temporal Convolutional Network for Direct Auditory Attention Decoding

Authors: Iñigo García-Ugarte, Rubén Eguinoa, Ricardo San Martín, Daniel Paternain, Carmen Vidaurre

Organizations: Department of Science, Universidad Pública de Navarra (UPNA), 31006 Pamplona, Spain · BCBL, Basque Center on Cognition Brain and Language, 20009 San Sebastián, Spain · Department of Statistics, Computer Sciences and Mathematics, Universidad Pública de Navarra (UPNA), 31006 Pamplona, Spain · Ikerbasque, Basque Foundation for Science, 48009 Bilbao, Spain ·

Abstract

A promising approach for steering auditory attention in complex listening environments relies on Auditory Attention Decoding (AAD), which aim to identify the attended speech stream in a multiple speaker scenario from neural recordings. Entrainment-based AAD approaches, typically assume access to clean speech sources and electroencephalography (EEG) signals to exploit low-frequency correlations between the neural response and the attended stimulus. In this study, we propose CA-TCN, a Causal-Anticausal Temporal Convolutional Network that directly classifies the attended speaker. The proposed architecture integrates several best practices from convolutional neural networks in sequence processing tasks. Importantly, it explicitly aligns auditory stimuli and neural responses by employing separate causal and anticausal convolutions respectively, with distinct receptive fields operating in opposite temporal directions. Experimental results, obtained through comparisons with three baseline AAD models, demonstrated that CA-TCN consistently improved decoding accuracy across datasets and decision windows, with gains ranging from 0.5% to 3.2% for subject-independent models and from 0.8% to 2.9% for subject-specific models compared with the next best-performing model, AADNet. Moreover, these improvements were statistically significant in four of the six evaluated settings when comparing Minimum Expected Switch Duration distributions. Beyond accuracy, the model demonstrated spatial robustness across different conditions, as the EEG spatial filters exhibited stable patterns across datasets. Overall, this work introduces an accurate and unified AAD model that outperforms existing methods while considering practical benefits for online processing scenarios. These findings contribute to advancing the state of AAD and its applicability in real-world systems.

Figures & tables

Fig. 1: (a) CA-TCN block diagram consisting of a Spatial Projection layer, a Temporal Convolutional Network (TCN) module, and a Classification module. EEG and audio signals are processed separately through two independent branches until classification module. (b) Design of the one-dimensional convolutional block within TCN.

Fig. 2: Illustration of dilated convolutions in a three-layer TCN (

N=3

) with exponential dilation factors

d=[1,2,4]

. The yellow sample (decision point) depends on the blue samples conforming its receptive field. Convolutions are applied causally for the stimulus (b) and anticausally for the EEG (a), with zero-padding on the left or right, respectively.

Dataset	No. of Subjects	No. of Trials	Trial Length	No. of Electrodes	Experimental Conditions
Jaulab	20	92	26 s	61	Loudspeakers, $360^{\circ}$ Fixed and moving auditory scenarios
DTU	18	60	50 s	64	Insert earphones, $\pm 60^{\circ}$ 3 different acoustic conditions
KULeuven	16	20	6 min. first 8 trials 2 min. last 12 trials	64	Insert earphones, $\pm 90^{\circ}$ Head-Related Transfer Function and Dichotic audio

TABLE I: Brief information about the data employed in the study

Model State	Cx	Cs	Nx	Ns	K	Parameters	MACs	Accuracy (5s)
Raw Input	✕	✕	✕	✕	✕	0.1K	0.1K	53.9%
+ Spatial Projection	32	32	✕	✕	✕	4.1K	2.1M	55.4%
+ TCN	32	32	5	5	3	27.1K	37.6M	71.8%
+ Causality	32	32	4	4	3	22.5K	30.5M	72.0%
+ Asymmetric RF (CA-TCN)	32	32	3	5	3	22.5K	32.9M	72.1%

TABLE II: Ablation study results obtained by progressively incorporating architectural components, starting from a correlation-and-classification-only baseline and culminating in the final optimized CA-TCN model. The accuracy corresponds with the validation accuracy obtained using a five second decision window for the subject-independent setting on the Jaulab dataset.

Dataset	Model	Accuracy (%)						MESD (s)
Dataset	Model	1s	2s	5s	10s	25s	50s	MESD (s)
Jaulab	Ridge	54.2 $\pm$ 1.9	56.0 $\pm$ 2.7	59.6 $\pm$ 3.9	63.5 $\pm$ 5.4	69.1 $\pm$ 7.4	74.6 $\pm$ 9.8	$>100$
	CCA	55.1 $\pm$ 1.9	58.3 $\pm$ 3.2	63.3 $\pm$ 5.0	68.3 $\pm$ 7.1	75.4 $\pm$ 9.6	81.4 $\pm$ 11.5	$>100$
	AADNet	57.5 $\pm$ 2.6	61.6 $\pm$ 3.6	68.6 $\pm$ 5.1	74.5 $\pm$ 6.5	82.3 $\pm$ 7.7	86.4 $\pm$ 8.6	26.9 $\pm$ 16.7
	CA-TCN	58.0 $\pm$ 2.4	62.5 $\pm$ 4.0	69.6 $\pm$ 5.6	75.4 $\pm$ 7.1	84.0 $\pm$ 8.8	88.5 $\pm$ 7.7	25.1 $\pm$ 20.8
DTU	Ridge	56.0 $\pm$ 2.8	60.1 $\pm$ 4.2	65.6 $\pm$ 6.4	70.2 $\pm$ 7.9	78.1 $\pm$ 10.1	83.6 $\pm$ 11.2	82.4 $\pm$ 129.8

TABLE III: Model performance comparison for Subject-Independent validation across three datasets

Fig. 3: Subject-independent (SI) distribution of MESD across the three datasets considered in this study. Each point corresponds to the MESD value obtained for a given subject, and median values are highlighted in the distributions. The number of subjects whose MESD exceeds 80 s are indicated at the top of the distributions in parentheses. Statistical analysis: ***: $p<0.001$ , **: $0.001\leq p<0.01$ , *: $0.01\leq p<0.05$ , None: $p\geq 0.05$ .

Dataset	Model	Accuracy (%)						MESD (s)
Dataset	Model	1s	2s	5s	10s	25s	50s	MESD (s)
Jaulab	Ridge	56.7 $\pm$ 2.7	60.7 $\pm$ 4.0	67.2 $\pm$ 6.0	73.5 $\pm$ 7.5	82.2 $\pm$ 8.4	89.0 $\pm$ 8.6	58.7 $\pm$ 109.4
	CCA	57.5 $\pm$ 2.9	62.1 $\pm$ 4.7	69.2 $\pm$ 6.9	75.5 $\pm$ 8.2	85.1 $\pm$ 8.9	92.3 $\pm$ 7.9	78.1 $\pm$ 205.1
	AADNet	59.4 $\pm$ 2.9	64.8 $\pm$ 4.8	73.0 $\pm$ 6.7	79.1 $\pm$ 8.1	86.3 $\pm$ 8.0	89.7 $\pm$ 7.1	20.1 $\pm$ 17.2
	CA-TCN	60.4 $\pm$ 3.1	66.1 $\pm$ 4.5	74.3 $\pm$ 6.3	81.3 $\pm$ 7.0	89.2 $\pm$ 6.2	91.6 $\pm$ 4.5	15.4 $\pm$ 10.4
DTU	Ridge	58.8 $\pm$ 2.1	65.4 $\pm$ 3.2	73.3 $\pm$ 4.7	79.8 $\pm$ 5.4	87.9 $\pm$ 5.2	94.0 $\pm$ 4.3	16.6 $\pm$ 7.6

TABLE IV: Model performance comparison for Subject-Specific validation across three datasets

Fig. 4: Subject-Specific (SS) distribution of MESD across the three datasets considered in this study. Each point corresponds to the average MESD obtained for a given subject, and median values are highlighted in the distributions.

Fig. 5: Topographic maps of the first cluster obtained by grouping the Spatial Projection filters from the CA-TCN model across the Jaulab (a), DTU (b), and KULeuven (c) datasets. The values below the topographic maps indicate the cosine similarity between each pair of clusters. These maps reveal a predominance of fronto-central, left-lateralized areas.

Neural Decoding

Neural decoding maps brain activity or error-correction syndromes to stimuli, behavior, language, or corrective actions.

See all 179 papers

AFA-Net: A Differential Attention Approach for Auditory Attention Detection
Sep 25, 2026Philip H. Lee, Shreeram Suresh Chandra, Karan Thakkar +1Differential Attention EEG Decoding
Independent Researcher, USA · Center for Language and Speech Processing (CLSP), Johns Hopkins University, USA · Laboratory for Computational Audio Perception (LCAP), Johns Hopkins University, USA +1
End-to-End Markov State Sequence Learning for Auditory Attention Decoding
Jul 21, 2026Yushan Yashengjiang, Jie Zhang, Miao Sun +3Markov Models Neural Decoding
NERC-SLIP, University of Science and Technology of China (USTC), Hefei 230027, China · School of Information and Communication Engineering, Guangzhou Maritime University, Guangzhou 510725, China · Artificial Intelligence Research Institute, iFLYTEK Company, Ltd., Hefei 230088, China +1
FAConformer: Frequency-Aware Convolutional Transformer for Auditory Attention Decoding
Jun 12, 2026Ziwei Wang, Xingyi He, Tianwang Jia +2Hybrid CNN-Transformer Architectures Neural Decoding
Hubei Key Laboratory of Brain-inspired Intelligent Systems, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China

Sep 25, 2026J/K move · Enter open · S save

AFA-Net: A Differential Attention Approach for Auditory Attention Detection

Philip H. Lee, Shreeram Suresh Chandra, Karan Thakkar, John H. L. Hansen

Independent Researcher, USA · Center for Language and Speech Processing (CLSP), Johns Hopkins University, USA · Laboratory for Computational Audio Perception (LCAP), Johns Hopkins University, USA+1

Differential Attention EEG Decoding

Open paper AskPDF arXiv

Auditory Attention Detection (AAD) utilizes electroencephalographic (EEG) signals to identify a target speaker in a multi-speaker environment. Despite considerable progress, existing deep learning architectures often lack explicit mechanisms for handling noisy EEG data. To address this limitation, we propose Auditory Focus Attention Networks (AFA-Net), a machine learning framework that replaces vanilla attention with a simple yet flexible differential attention mechanism to help focus on task-relevant neural activity. AFA-Net achieves an upward accuracy of 96.8% at the 2s decision window, while using substantially fewer parameters than most existing methods. To the best of our knowledge, AFA-Net is among the first frameworks to explicitly try to combat EEG noise to improve AAD.

Abstract

Neural Decoding

Explore similar work

AFA-Net: A Differential Attention Approach for Auditory Attention Detection

End-to-End Markov State Sequence Learning for Auditory Attention Decoding

FAConformer: Frequency-Aware Convolutional Transformer for Auditory Attention Decoding