Paper deep dive
Addressing Limited Data in Auditory Attention Decoding with Diffusion Generative Models
David Rannaleet, Victor Gunnarsson, Bo Bernhardsson, Martin A. Skoglund, Emina Alickovic
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/22/2026, 2:25:29 AM
Summary
This paper investigates the use of Diffusion Probabilistic Models (DPMs) to generate synthetic speech-evoked EEG data to address data scarcity in Auditory Attention Decoding (AAD) for hearing aids. The authors evaluate DPM-based data augmentation for Locus-of-Attention (LoA) classification tasks, demonstrating that incorporating synthetic EEG data significantly improves AAD performance compared to models trained solely on measured EEG data.
Entities (9)
Relation Signals (6)
Diffusion Probabilistic Models → usedfor → Synthetic EEG Generation
confidence 95% · we investigate diffusion probabilistic models (DPMs) for generating synthetic speech-evoked EEG data.
Auditory Attention Decoding → appliedin → Hearing Aids
confidence 94% · AAD uses electroencephalogram (EEG) data to decode listener's attention, enabling real-time tracking of specific sound sources in hearing aids.
Synthetic EEG Data → usedforaugmentation → Auditory Attention Decoding
confidence 93% · incorporating synthetic data significantly improves AAD performance compared to models trained solely on measured EEG data
Denoising Diffusion Implicit Models → istypeof → Diffusion Probabilistic Models
confidence 92% · We specifically selected Denoising Diffusion Implicit Models (DDIMs) ... as a class of latent variable generative models
U-Net → usedin → Diffusion Probabilistic Models
confidence 90% · We use a variant of the original U-Net architecture ... The U-Net takes a noisy EEG sample and corresponding diffusion time step as inputs
EEGNeX → usedfor → Locus-of-Attention Classification
confidence 88% · For LoA classification, we employ the EEGNeX classifier
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Limited training data constrains deep learning models for Auditory Attention Decoding (AAD) in hearing aids (HAs). AAD uses electroencephalogram (EEG) data to decode listener's attention, enabling real-time tracking of specific sound sources. However, achieving high AAD performance with short time windows typical in HAs (<=1s) is challenging due to the scarcity of real-world speech-evoked EEG data. To address this issue, we investigate diffusion probabilistic models (DPMs) for generating synthetic speech-evoked EEG data. DPMs learn the underlying complex data structure through a denoising process and can generate realistic samples suitable for data augmentation. We evaluate the use of synthetic EEG data for augmenting datasets in locus-of-attention (LoA) classification tasks. Our experiments demonstrate that DPMs can generate realistic EEG signals and that incorporating synthetic data significantly improves AAD performance compared to models trained solely on measured EEG data (p<0.05). These results highlight the potential of diffusion-based data augmentation to mitigate training data limitations and improve the robustness of short-window AAD models in HA applications.
Tags
Links
- Source: https://arxiv.org/abs/2607.18345v1
- Canonical: https://arxiv.org/abs/2607.18345v1
Trouble viewing inline? Open PDF directly →
Full Text
30,888 characters extracted from source content.
Expand or collapse full text
Addressing Limited Data in Auditory Attention Decoding with Diffusion Generative Models* David Rannaleet 1,† , Victor Gunnarsson 1,† , Bo Bernhardsson 1 , Martin A. Skoglund 2,3,‡ , and Emina Alickovic 2,3,‡ Abstract—Limited training data constrains deep learning mod- els for Auditory Attention Decoding (AAD) in hearing aids (HAs). AAD uses electroencephalogram (EEG) data to decode listener’s attention, enabling real-time tracking of specific sound sources. However, achieving high AAD performance with short time windows typical in HAs (≤ 1s) is challenging due to the scarcity of real-world speech-evoked EEG data. To address this issue, we investigate diffusion probabilistic models (DPMs) for generating synthetic speech-evoked EEG data. DPMs learn the underlying complex data structure through a denoising process and can generate realistic samples suitable for data augmentation. We evaluate the use of synthetic EEG data for augmenting datasets in locus-of-attention (LoA) classification tasks. Our experiments demonstrate that DPMs can generate realistic EEG signals and that incorporating synthetic data significantly improves AAD performance compared to models trained solely on measured EEG data (p < .05). These results highlight the potential of diffusion-based data augmentation to mitigate training data limitations and improve the robustness of short-window AAD models in HA applications. Index Terms—Auditory Attention Decoding, EEG, Speech, Generative Models, Diffusion Models I. INTRODUCTION Deep learning has advanced hearing aids (HAs), aiming to improve performance in challenging listening environments with competing sound sources [2]. However, existing HA amplification strategies often fall short in such scenarios. Auditory attention decoding (AAD) emerges as a promising personalized solution [3]–[5]. AAD uses brain responses to speech, recorded with an electroencephalogram (EEG), to decode the listener’s attention, potentially enabling real-time tracking of desired sound sources in HAs. Despite recent advances demonstrating the effectiveness of EEG-based AAD and its potential for non-intrusive, real-time HA enhancement, a key challenge remains: achieving reliable decoding with short time windows (one second or less) typical for HAs. This limitation is largely due to insufficient training data for deep learning models. Speech-evoked EEG recordings are often limited in duration, as most studies rely on short sessions This paper’s data analysis is based on a Master’s Thesis work [1] Thanks to the ELLIIT strategic research programme for funding. 1 Department of Automatic Control, Lund University, Lund, Sweden. bob at lth.lu.se 2 EriksholmResearchCentre,Snekkersten,Denmark. mnsk,eali,dowe at eriksholm.com 3 Department of Electrical Engineering, Link ̈ oping University, Link ̈ oping, Sweden. martin.skoglund,emina.alickovic at liu.se † Equally contributed as first authors. ‡ Equally contributed as last authors. per subject, and larger, more diverse datasets are needed to improve generalization across listening conditions. Generative learning provides a partial solution to this data scarcity challenge by learning patterns in speech-evoked EEG and generating realistic, task-specific signals. Such synthetic data can augment existing datasets, improving AAD perfor- mance when real-world data collection is constrained. Diffusion probabilistic models (DPMs), a class of latent variable generative models, are particularly well suited for this task. By learning to reverse a gradual noise-adding process, DPMs generate realistic samples [6]–[8]. Their success in im- age synthesis and natural language processing suggests strong potential for EEG-based AAD. In this context, DPMs offer two main advantages: they can model complex, non-linear relationships between EEG and auditory stimuli, and they can learn task-specific EEG patterns associated with attention. Although diffusion probabilistic models have previously been applied to EEG-based tasks such as sleep stage classification and cognitive decline detection using long analysis windows (e.g., 30s) [9], [10], real-time auditory attention decoding requires reliable inference from short time windows (≤ 1s), motivating generative models capable of capturing informative short-term EEG representations. This work addresses limited EEG data for training AAD models by focusing on attended-talker classification in multi-talker environments. We consider locus-of-attention (LoA) classification, which decodes attended speech direc- tion (left vs. right), and propose a DPM-based approach to generate synthetic speech-evoked EEG data. We evaluate DPM-based data augmentation for LoA tasks and demonstrate improved AAD performance, supporting more accurate and robust real-time attention decoding in future HAs I. METHODOLOGY A. Experimental Design a) Dataset: This study uses an EEG data set previously analyzed in [1], [2], [11], [12] using different methods. The dataset comprises recordings from 34 participants (24 male) aged 21-84 (mean: 64.2 years, SD: 13.6 years) with symmetri- cal sensorineural hearing loss (average 4-frequency pure-tone audiometry of 47.5 dB). All participants were experienced HA users with no reported neurological disorders, dyslexia, or diabetes. The study was approved by the ethics committee of the Capital Region of Denmark (journal no. H-1-2011-033), and written informed consent was obtained. arXiv:2607.18345v1 [cs.SD] 20 Jul 2026 b) Recordings: EEG data were recorded at 1024 Hz using the BioSemi Active Two recording system with 64 electrodes (international 10-20 system) and two mastoid ref- erence electrodes. Participant were seated in a sound-proofed booth facing a screen. The experimental setup consisted of six loudspeakers arranged around the participant, positioned at ±30 ◦ in front, ±112.5 ◦ laterally, and ±157.5 ◦ behind. c) Stimuli: The experiment involved a two-talker lis- tening task with background noise. Danish news clips of neutral content spoken by male and female speakers were presented from the front loudspeakers, while background noise simulating a 16-talker environment was presented from rear loudspeakers at -3dB relative to the foreground speech. d) Trials: Participants completed 80 experimental trials (excluding familiarization trials) across four sessions using a 2×2 hearing-aid configuration, in which two different noise reduction schemes (NR) were each evaluated with NR enabled or disabled. Each trial comprised a 5-second background-noise onset followed by 33 seconds of concurrent target speech and background noise, yielding a total duration of 38 seconds. Within each session, trials were organized into blocks of five with fixed target speaker direction and gender, while speech content was randomized across trials. Participants were instructed to attend to the target speaker and completed a content-related two-choice question after each trial. e) Preprocessing: Trials were preprocessed to remove noisy and irrelevant (ocular, muscular, and cardiac) component using band-pass (0.5–70Hz) and notch (49–51Hz) filtering, downsampling from 1024Hz to 256Hz, visual inspection with channel interpolation, and independent component analysis [13]. Data from 31 participants remained after excluding recordings with persistent artifacts. f) Data Split and Normalization: The final dataset com- prises 2420 trials of 33 s, segmented into 1 s EEG samples using a sliding window, yielding a balanced set of left- and right-target labels (1210 each). Data were split into training (60%), validation (20%), and testing (20%) sets by assigning the first three trials within each block to training, the fourth to validation, and the fifth to testing, ensuring balanced class distributions across splits [12]. EEG data were standardized using the mean and standard deviation computed from the training set across all channels. B. EEG Diffusion Modeling DPMs provide an effective approach to data augmentation by gradually adding noise to the data during a forward diffusion process and learning to reverse it through denoising. In forward diffusion, noise is gradually added by progressing through Markov chain via multiplication with the Markov transition kernel q(x 1:T |x 0 ) = Q T t=1 q(x t |x t−1 ) with t = 0 denoting the original data distribution q(x 0 ) and t = T is the final instance of the chain representing the converted data distribution q(x t ). The reverse process, as defined by [6], is p θ (x 0:T ) = p(x T ) Q T t=1 p θ (x t−1 |x t ), where p θ (x t−1 |x t ) is the reverse Markov transition kernel. The Gaussian forward kernel is defined as q(x t |x t−1 ) = N (x t ; √ 1− β t x t−1 ,β t I ), where β t is the variance of the introduced noise. Given a small size of the variance β t , the reverse Markov transition kernel will be of the same functional form as the forward process: p θ (x t−1 |x t ) =N (x t−1 ;μ θ (x t ,t), Σ θ (x t ,t)). Here, the mean μ θ (x t ,t) and covariance Σ θ (x t ,t) are unknown parameters, which are estimated by a neural network. a) ImplicitDiffusion:Wespecificallyselected Denoising Diffusion Implicit Models (DDIMs) [14] due to their ability to generate new samples without requiring sequential time steps, unlike e.g. Denoising Diffusion Probabilistic Model (DDPM). In DDIM, we use the forward distributions indexed by σ, q σ (x 1:T |x t ,x 0 ) = q σ (x T |x 0 ) Q T t=2 q σ (x t−1 |x t ,x 0 ) where q σ (x t−1 |x t ,x 0 ) = N √ ̄α t−1 x 0 + p 1− ̄α t−1 − σ 2 t · x t − √ ̄α t x 0 / p 1− ̄α t−1 ,σ 2 t I , with α and its cumulative product as ̄α, α t = 1 − β t and ̄α t = Q t s=1 α s . The DDIM training loop is identical to that of DDPM, as it models the same parameter ε θ and uses the same forward diffusion for adding noise. Since the reverse diffusion process is determined by the forward diffusion process, it is possible to use another forward diffusion q σ defined on a subset of latent variables x τ 1 ,...,x τ S during sampling, thus reducing the number of steps from T to S as x τ i−1 (η)= √ ̄α τ i−1 x τ i − p 1− ̄α τ i ε (τ i ) θ (x τ i ) / p ̄α τ i + p 1− ̄α τ i−1 − σ τ i (η) 2 ε (τ i ) θ (x τ i )+ σ τ i (η)ε i where σ τ i (η) = η p 1− ̄α τ i−1 /1− ̄α τ i p 1− ̄α τ i / ̄α τ i−1 and η is a hyperparameter that sets the stochasticity of the process. The variances (β) of the diffusion model are set when initializing the DDIM schedule using a β scheduler. Our implementation uses the squared cosine scheduler proposed by [15]. Each β t is now set to the value of β t = 1 − ̄α t / ̄α t−1 , where ̄α t = f (t) /f (0),f (t) = cos ( t/T + s /1 + s· π /2) 2 . The initial value x 0 ∼ x τ i − p 1− ̄α τ i ε (τ i ) θ (x τ i ) / p ̄α τ i is updated for each step in the sampling process. b) U-Net: We use a variant of the original U-Net ar- chitecture [16], commonly used for image domain diffusion. The U-Net takes a noisy EEG sample and corresponding diffusion time step as inputs, and outputs a of the mean μ θ (x t ,t) in the reverse diffusion kernel p θ (x t−1 |x t ). The implementation is adapted from the HuggingFace diffusers library, built on transformers library [17]. Our U-Net uses a base of six convolution blocks each for the encoder and decoder. All convolutions in the network use a 3× 3 kernel and the Sigmoid-Linear Unit (SiLU) activation function. To provide time-step information absent in the original U-Net, we incorporate sinusoidal positional embeddings, following the DDPM formulation [7]. c) V-prediction: Diffusion models are commonly trained to predict the added noiseε, referred to here as the Epsilon loss. Alternatively, the model can predict the sample (x) itself or a combination of both, known as v-prediction (V-pred) [18]. In V-pred, a new prediction target v is defined as v ≡ α t ε− σ t x with the corresponding loss as ||v− ˆv|| 2 . d) Spectral Loss: To consider spectral components of EEG, a spectral loss using the Short-Time Fourier Transform (STFT) was used, which takes the form of: 1 /N∥|STFT (x)|− |STFT ( ˆ x)|∥ 2 2 . Here, x represents the original data, ˆ x is the predicted data and N is the number of elements in x. This approach is inspired by the spectral loss used as a reconstruction loss of a VQ-VAE in [19] to consider mid- and high frequencies. The data is scaled such that the loss falls within the interval [0, 0.5] to balance it with the existing average MSE loss. e) Jensen-Shannon Distance: To quantitatively assess the similarity between real EEG and generated EEG data distributions, the Jensen-Shannon Distance (JSD) was used. It is defined as the square-root of the Jensen-Shannon divergence 1 /2· D KL (P∥ P + Q /2) + 1 /2· D KL (Q∥ Q + P /2), where P and Q represent two distributions being compared, and D KL (·) is the Kullback-Liebler (KL) divergence. f) Implementation: To reduce sensitivity to STFT pa- rameter choices, the mean is taken across multiple STFT resolutions, preventing overfitting to one STFT representation and allowing the model to capture a broader range of time- frequency structure [20]. Loss implementations follow a mod- ified version of the code from [21]. Training uses T = 1000 diffusion steps [6], [7]. Input samples are one-second EEG segments with a shape of [N ,1,64,256] and a batch size of N = 64. The model is trained for 100 epochs using the AdamW optimizer [22], an initial learning rate of 1e-4 and a cosine schedule with 500 warmup steps. In our testing, 100 epochs were sufficient for loss convergence. To prevent sample saturation, dynamic thresholding is ap- plied [23]. At each diffusion step a percentile-based threshold s is computed and values beyond ±s are clipped. Because the EEG data are standardized rather than normalized, we cap the threshold at s max = 5 (i.e., if s > s max , we use s max ) and do not rescale after clipping, thereby bounding values to [−5, 5] and suppressing extreme outliers. Separate diffusion models are trained for left- and right-attention labels, and 45,000 one- second EEG samples are generated per label. Full training details are provided in our GitHub repository. 1 C. Locus of Attention Classification AAD aims to identify the attended speaker in multi-speaker environments, with locus-of-attention (LoA) methods forming a subset that decode the attended speech direction (left vs. right) from EEG [12], [24]–[26]. For LoA classification, we employ the EEGNeX classifier [27], selected for its strong per- formance across EEG classification tasks compared to earlier convolutional neural network models such as EEGNet [28]. We use a PyTorch reimplementation of the original EEGNeX architecture [29]. Classifiers were trained with a batch size of 64 for 100 epochs using the AdamW optimizer [22] and a fixed learning rate of 5 × 10 −4 . EEG data were sampled using 1-second sliding windows with a stride of 0.75 s, resulting in a 25% overlap between consecutive samples and no overlap across trials. Data were split into training (60%), testing (20%), and validation (20%) sets, and for each configuration, the model 1 Link provided upon manuscript decision. with the lowest test-set loss was evaluated on the held-out validation set. To assess the impact of synthetic EEG data for data aug- mentation, we first trained a baseline classifier and reused its hyperparameters across all models for consistency. Augmented models combined real EEG training data with synthetic EEG at different ratios (15%, 30%, 60%, or 100% of the original training set size), and each configuration was trained 20 times to estimate average performance and confidence intervals. Since diffusion models generate data from normally distributed noise, we included a noise addition model as a simpler augmentation baseline, in which Gaussian noise sampled from N (0, 0.15) was added to each data point of each EEG channel. The standard deviation of the noise was empirically tuned by visually inspecting samples before and after noise addition so that the original signal remained identifiable. This noise addition was applied to 15% of the training data. I. RESULTS AND DISCUSSION A. Diffusion Model Performance Directly computing JSD on high-dimensional EEG data is computationally demanding; therefore, EEG distributions were approximated using channel-wise histograms. For each atten- tion label (Left or Right), 15,000 random samples were drawn from standardized real and generated EEG data. Per-channel histograms were computed using 200 equal-width bins over the range [−10, 10], and the resulting distributions (P and Q) were used to compute JSD. Across diffusion model losses (Epsilon, V-pred, and Spec- tral) and attention targets, JSD values were comparable. Ep- silon showed JSD values of 0.026 (Left) and 0.028 (Right); V-pred showed higher JSD values (0.064 and 0.072); and Spectral had intermediate JSD values of 0.042 (Left) and 0.047 (Right). These results indicate limited dependence of distributional similarity on the specific diffusion configuration, with Epsilon and Spectral models showing closely matched similarity to real EEG data. Figure 1 shows channel-wise JSD for the Epsilon and Spectral models. Two specific channels are highlighted with corresponding histograms. Most channels exhibit low diver- gence, with fronto-central regions showing particularly high similarity between real and synthetic EEG. Most channels show minimal JSD variations, making the differences visually subtle; slight deviations occur near zero, where real EEG displays higher density. This spatial pattern corresponds to known attention-related cortical activity [30]–[32], suggesting that the generative models capture task-relevant EEG structure. B. Classification Figure 2 compares LoA classification accuracy between the baseline, the ’noise addition’ model, and the diffusion models. The boxplot shows the interquartile range containing 50 % of the data, with a median marked by a line. Data outside the whiskers are considered outliers. The ’noise addition’ model shows a drop of around 3% relative to the baseline, indicating that adding noise alone does not improve performance. Most Fig. 1: JSD between real and synthetic EEG for two models. The topoplot shows channel-wise JSD, and two channels are highlighted with corresponding histograms (red: real, blue: synthetic). The largest differences appear near zero. diffusion models perform similarly to the baseline, with some achieving modest but statistically supported improvements. Fig. 2: Box-plot of LoA classification accuracy across models. Each dot shows a single run; circled points indicate outliers beyond ±1.5 times the interquartile range. To assess statistical differences relative to the baseline, we applied Dunnett’s test [33]. Table I summarizes the resulting p-values. Three configurations differed from the baseline: the noise-addition model showed significantly lower accuracy (p < 10 −4 ), whereas the Epsilon model with 60% synthetic data (p = 0.083) and the Spectral model with 100% synthetic data (p = 0.042) achieved higher accuracy. TABLE I: P-values from Dunnett’s test vs. baseline (n = 20). ≪ 0.01 denotes p < 10 −4 . Model15%30%60%100% Epsilon (LoA)1.00.180.0830.12 Spectral (LoA)1.01.00.980.042 V-pred (LoA)1.01.01.00.61 Noise Addition ≪ 0.01 (15% only) Table I reports the accuracy improvements for the two diffusion-based configurations with p < 0.10. The Epsilon model with 60% synthetic data improves accuracy by 0.71%, and the Spectral model with 100% synthetic data improves accuracy by 0.78%. Although the confidence intervals overlap, indicating no clear difference between the two configurations, both outperform the baseline. TABLE I: Mean improvement (MI, n = 20) in accuracy (Acc) for the models with significant P-values (< 0.10). The confidence interval (CI) refers to the range of MI. ModelAdded data (%)Acc (%)MI (%)95% CI Baseline073.05N/AN/A Epsilon6073.760.710.12 to 1.31 Spectral10073.830.780.16 to 1.41 This study relies on a single EEG dataset, which limits generalizability. However, our methods do not depend on dataset-specific features, suggesting potential applicability to other datasets. Future work should evaluate additional EEG corpora and recording conditions. The analysis also assumes that trial-level labels accurately reflect attention and that each 1 s segment contains sufficient information about the attended talker. Occasional label mismatches are possible and may affect training and evaluation. Computational constraints limited the number of classifier configurations explored, and only 20 runs per setting were per- formed for statistical comparisons. Although diffusion models are computationally intensive to train, once trained they gen- erate synthetic EEG much faster than acquiring equivalent real data, partially mitigating this limitation. We did not assess whether diffusion models may reproduce near-duplicate EEG samples, a phenomenon observed in other diffusion domains [34]. Duplication could introduce bias or reduce generalization and should be examined in future work. Finally, as synthetic EEG becomes more realistic, safeguards may be needed to prevent generated signals from being misin- terpreted as clinical data, and clear usage guidelines may help mitigate this risk. IV. CONCLUSION This work assessed the use of DPMs for generating syn- thetic speech-evoked EEG to support AAD. The generated signals improved classification of attended-speech direction compared to models trained on measured EEG alone (p < .1), indicating that DPM-based augmentation can help address data limitations in AAD. Several directions remain for future work. Further tuning of DPM parameters is needed, as this was not addressed in detail. Replacing the U-Net neural network architecture with a model tailored for multichannel time-series data could be a promising direction. While U-Net is effective for image-based diffusion models, it is likely not optimally suited for EEG data. Developing an inner denoising model that better differentiates noise from distinctive EEG features could improve generation quality. Such a model would need to effectively capture the temporal and spatial dynamics inherent in EEG signals. Furthermore, exploring conditional and multimodal diffu- sion models might improve the realism and task-specificity of the generated EEG data. Developing subject-specific models that can adapt to individual EEG patterns is another possible direction for more personalized AAD applications. Finally, adapting the diffusion model to handle continuous EEG data of arbitrary lengths, both as input and output, could provide much greater task flexibility. This approach would enable the generation of longer, more coherent sequences that better mimic real EEG signals. DECLARATION OF GENERATIVE AI AND AI-ASSISTED TECHNOLOGIES During the preparation of this work, the authors used ChatGPT-5.1 to improve readability through rephrasing and grammar checking. The authors reviewed and edited the con- tent as required and take full responsibility for the publication. REFERENCES [1] Rannaleet, David and Gunnarsson, Victor, “Diffusion Modelling ap- proaches to EEG-based Auditory Attention Decoding,” 2024, Master’s Thesis, Lund University. [2] A. H. Andersen, S. Santurette, M. S. Pedersen, E. Alickovic, L. Fiedler, J. Jensen, and T. Behrens, “Creating clarity in noisy environments by using deep learning in hearing aids.” Seminars in Hearing, vol. 42, no. 3, p. 260 – 281, 2021. [3] J. A. O’sullivan, A. J. Power, N. Mesgarani, S. Rajaram, J. J. Foxe, B. G. Shinn-Cunningham, and et.al., “Attentional selection in a cocktail party environment can be decoded from single-trial EEG,” Cerebral cortex, vol. 25, no. 7, p. 1697–1706, 2015. [4] E. Alickovic, T. Lunner, F. Gustafsson, and L. Ljung, “A tutorial on auditory attention identification methods.” Frontiers in Neuroscience, vol. 13, 2019. [5] S. Geirnaert, S. Vandecappelle, E. Alickovic, A. de Cheveigne, E. Lalor, B. T. Meyer, S. Miran, T. Francart, and A. Bertrand, “Electroencephalography-based auditory attention decoding: Toward neurosteered hearing devices.” IEEE Signal Processing Magazine, vol. 38, no. 4, p. 89 – 102, 2021. [6] J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning.PMLR, 2015, p. 2256–2265. [7] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, Inc., 2020, p. 6840–6851. [8] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” arXiv preprint arXiv:2011.13456, 2020. [9] B. Aristimunha, R. Y. de Camargo, S. Chevallier, O. Lucena, A. G. Thomas, M. J. Cardoso, W. H. L. Pinaya, and J. Dafflon, “Synthetic Sleep EEG Signal Generation using Latent Diffusion Models,” in Deep Generative Models for Health Workshop NeurIPS 2023, 2023. [10] G. Sharma, A. Dhall, and R. Subramanian, “MEDiC: Mitigating EEG Data Scarcity Via Class-Conditioned Diffusion Model,” in Deep Gener- ative Models for Health Workshop NeurIPS 2023, 2023. [11] E. Alickovic, E. H. N. Ng, L. Fiedler, S. Santurette, H. Innes-Brown, and C. Graversen, “Effects of hearing aid noise reduction on early and late cortical representations of competing talkers in noise,” Frontiers in Neuroscience, vol. 15, 2021. [12] M. A. Tanveer, M. A. Skoglund, B. Bernhardsson, and E. Alickovic, “Deep learning-based auditory attention decoding in listeners with hearing impairment,” Journal of Neural Engineering, 2024. [13] O. Keding, J. Wilroth, M. A. Skoglund, and E. Alickovic, “Effect of independent component artifact rejection on eeg-based auditory atten- tion decoding,” in 2024 32nd European Signal Processing Conference (EUSIPCO). IEEE, 2024, p. 877–881. [14] J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. [15] A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilis- tic models,” in International conference on machine learning. PMLR, 2021, p. 8162–8171. [16] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Horneg- ger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer International Publishing, 2015, p. 234–241. [17] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, and et.al., “Transformers: State-of-the-art natural language processing,” in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Q. Liu and D. Schlangen, Eds. Association for Computational Linguistics, Oct. 2020, p. 38–45. [18] T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in International Conference on Learning Represen- tations, 2022. [19] P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, “Jukebox: A generative model for music,” arXiv preprint arXiv:2005.00341, 2020. [20] R. Yamamoto, E. Song, and J. Kim, “Parallel WaveGAN: A Fast Wave- form Generation Model Based on Generative Adversarial Networks with Multi-Resolution Spectrogram,” in 2020 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2020, Barcelona, Spain, May 4-8, 2020. IEEE, 2020, p. 6199–6203. [21] C. J. Steinmetz and J. D. Reiss, “auraloss: Audio focused loss functions in PyTorch,” in Digital Music Research Network One-day Workshop (DMRN+15), 2020. [22] F. Hutter and I. Loshchilov, “Decoupled weight decay regularization,” in International Conference on Learning Representations (ICLR), vol. 7, 2019. [23] C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al., “Photorealistic text-to-image diffusion models with deep language understanding,” Advances in neural information processing systems, vol. 35, p. 36 479–36 494, 2022. [24] S. Vandecappelle, L. Deckers, N. Das, A. H. Ansari, A. Bertrand, and T. Francart, “EEG-based detection of the locus of auditory attention with convolutional neural networks,” eLife, vol. 10, p. e56481, 4 2021. [25] J. Wilroth, B. Bernhardsson, F. Heskebeck, M. A. Skoglund, C. Bergeling, and E. Alickovic, “Improving EEG-based decoding of the locus of auditory attention through domain adaptation,” Journal of Neural Engineering, vol. 20, no. 6, p. 066022, 2023. [26] C. Puffay, B. Accou, L. Bollens, M. J. Monesi, J. Vanthornhout, H. V. hamme, and T. Francart, “Relating EEG to continuous speech using deep neural networks: a review,” Journal of Neural Engineering, vol. 20, no. 4, p. 041003, aug 2023. [27] X. Chen, X. Teng, H. Chen, Y. Pan, and P. Geyer, “Toward reliable signals decoding for electroencephalogram: A benchmark study to EEG- NeX,” Biomedical Signal Processing and Control, vol. 87, p. 105475, 2024. [28] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces,” Journal of Neural Engineering, vol. 15, no. 5, p. 056013, Jul. 2018. [29] J. Ansel, E. Yang, H. He, N. Gimelshein, A. Jain, M. Voznesensky, B. Bao, and et.al., “PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation,” in 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’24). ACM, Apr. 2024. [30] M. Giard, F. Perrin, J. Pernier, and F. Peronnet, “Several attention-related wave forms in auditory areas: a topographic study,” Electroencephalog- raphy and Clinical Neurophysiology, vol. 69, no. 4, p. 371–384, 1988. [31] M. J. Crosse, J. S. Butler, and E. C. Lalor, “Congruent visual speech enhances cortical entrainment to continuous auditory speech in noise- free conditions,” Journal of Neuroscience, vol. 35, no. 42, p. 14 195– 14 204, 2015. [32] O. Etard and T. Reichenbach, “Neural speech tracking in the theta and in the delta frequency band differentially encode clarity and comprehension of speech in noise,” Journal of Neuroscience, vol. 39, no. 29, p. 5750– 5759, 2019. [33] C. W. Dunnett, “A multiple comparison procedure for comparing several treatments with a control,” Journal of the American Statistical Associa- tion, vol. 50, no. 272, p. 1096–1121, 1955. [34] G. Somepalli, V. Singla, M. Goldblum, J. Geiping, and T. Goldstein, “Diffusion art or digital forgery? investigating data replication in diffu- sion models,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, p. 6048–6058.