blind source separation

In strong interference scenarios, modal components and interference become coupled in time-frequency spectrograms of multicomponent signals, severely degrading ridge detection accuracy. Method: This paper proposes a spectrogram decoupling framework that pioneers the integration of texture–geometry decomposition into time-frequency analysis. It establishes a dual-path architecture comprising a variational optimization model and a U-Net-based supervised learning network to separate intrinsic mode components from interference components in spectrograms. Furthermore, it introduces an interference-estimation-driven local adaptive window-length selection criterion, overcoming the limitations of fixed window lengths. Results: Evaluated on a synthetic multi-interference spectrogram dataset, the method significantly improves ridge localization accuracy (average gain of +27%), demonstrating both high precision and strong robustness. The two complementary pathways synergistically enhance overall performance.

Neural audio codecs often rely on data- or task-specific priors to disentangle frequency-band features, resulting in poor interpretability and limited generalizability. Method: We propose a generic soft disentanglement representation learning framework. It first applies spectral decomposition to project time-domain audio into orthogonal frequency-band subspaces; then employs a multi-branch encoder to model each band independently, jointly optimized via reconstruction and perceptual losses. Crucially, the framework imposes no assumptions about task structure or data distribution, enabling task-agnostic, soft intra-band semantic disentanglement. Contribution/Results: Experiments demonstrate significant improvements over state-of-the-art codecs in objective audio quality metrics (e.g., PESQ, STOI) and perceptual fidelity. Moreover, the learned representations exhibit strong generalization to downstream tasks—such as audio inpainting—and provide interpretable, structured frequency-band semantics without architectural or prior constraints.

Weak interpretability of representations and dataset- and task-specific disentanglement remain critical bottlenecks in neural audio codecs. To address this, we propose a time-domain neural codec based on spectral decomposition of the input signal, incorporating a soft frequency-band decoupling mechanism that explicitly models semantic independence across frequency subbands during encoding. Our approach further integrates spectral prior structures to guide representation learning. A differentiable separation loss enables joint time–frequency modeling, substantially improving semantic disentanglement and cross-task generalization. Experiments demonstrate that our model surpasses state-of-the-art baselines in reconstruction fidelity (STOI, PESQ) and perceptual quality (MOS). It exhibits strong robustness and versatility across diverse downstream tasks—including speech enhancement, source separation, and audio synthesis—without task-specific architectural modifications.

Addressing the long-standing challenge of single-channel nonlinear blind source separation (BSS), this paper proposes a novel multi-encoder autoencoder framework—the first to introduce a multi-encoder architecture for this task. The method centers on three key components: (i) an encoding mask mechanism that enforces specialization of latent representations into distinct source-specific subspaces; (ii) a sparse mixing loss that regularizes the mixing structure via sparsity-inducing priors; and (iii) a zero-reconstruction loss that ensures orthogonality and interpretability of recovered sources. The model is trained end-to-end to independently estimate and separate constituent sources from their nonlinear mixture. Extensive evaluation on synthetic data and real-world multichannel polysomnography (PSG) recordings—specifically electrocardiogram (ECG) and photoplethysmogram (PPG)-derived respiratory signals—demonstrates substantial performance gains over conventional single-channel BSS approaches. This work establishes a new paradigm for physiological signal disentanglement.

This study addresses the challenge of end-to-end reconstruction of high-fidelity, polyphonic, harmonically rich natural music directly from raw, non-invasive EEG signals. Methodologically, it introduces latent diffusion models (LDMs) to the EEG-to-audio decoding task for the first time, enabling direct mapping from raw temporal EEG to complex audio waveforms—without handcrafted features, channel selection, or signal preprocessing. A novel neural-embedding-based evaluation metric is proposed to better quantify auditory perceptual consistency. Experiments on the NMED-T dataset demonstrate substantial improvements in timbral and structural fidelity; reconstructed audio exhibits superior intelligibility and musicality compared to conventional linear models and VAE baselines. Key contributions include: (1) the first LDM framework explicitly designed for natural-music synthesis from EEG; (2) a full-brain, end-to-end decoding paradigm that maps raw EEG to high-quality audio; and (3) a semantics-aware evaluation framework tailored for neural decoding.

Deep learning–based EEG decoding often suffers from latent overfitting, severely impairing cross-scenario generalization and hindering real-world BCI deployment. To address this, we propose the weakly supervised Disentangled Decoding Decomposition (D3) framework. D3 leverages positional prediction within trial sequences to drive latent-space disentanglement; introduces a novel independent subnetwork architecture coupled with a nonlinear ICA–inspired disentanglement mechanism to isolate task-agnostic neural components; and establishes a cross-dataset component activation contrast paradigm that explicitly models linear separability in the latent space. Experiments demonstrate that D3 stably isolates physiologically meaningful components in motor imagery data while suppressing task-irrelevant artifacts. In sleep staging, D3 achieves significant few-shot accuracy gains—requiring only minimal labeled data—and generalizes robustly to real-world scenarios.

This work proposes a knowledge-driven, self-supervised approach to audio segmentation and source separation that circumvents the reliance on large-scale manually annotated data. By integrating external prior knowledge—such as musical scores—into the audio processing pipeline for the first time, the method leverages hidden Markov models to achieve effective segmentation and separation of music and film audio without requiring labeled training data. Evaluated on synthetic datasets, the approach demonstrates strong performance, and in real-world film soundtrack tests, it significantly outperforms purely data-driven methods when incorporating sound-class priors. This represents a notable advance toward annotation-free audio analysis through principled integration of domain knowledge.

Existing research on audio spoofing detection often overlooks real-world scenarios where speech coexists with environmental sounds and may be partially manipulated. To address this gap, this work introduces PC-Mix, the first dataset specifically designed for partial-component spoofing detection in mixed audio, featuring realistic scenarios with locally forged environmental sounds. We further propose a joint learning framework that operates across both speech and environmental sound components. Through a unified evaluation protocol, our experiments demonstrate that spoofing detection under mixed conditions is significantly more challenging, and models trained under target-matched conditions substantially outperform those directly transferred from single-component settings. This study fills a critical research void in partial spoofing detection involving environmental sounds and mixed acoustic conditions.

This work addresses the challenge of learning disentangled and interpretable latent representations from complex, non-stationary, high-dimensional time-varying signals, which exhibit rich time-frequency structures that conventional variational autoencoders (VAEs) struggle to model effectively. To this end, we propose the Decompositional Variational Autoencoder (DecVAE), which uniquely integrates signal decomposition priors directly into the VAE framework. DecVAE employs an encoder-only architecture and jointly leverages signal decomposition models, contrastive self-supervised tasks, and variational inference to learn multi-subspace latent representations aligned with the intrinsic time-frequency characteristics of the data. Extensive experiments on synthetic data and three scientific datasets demonstrate that DecVAE substantially outperforms existing VAE approaches, achieving significant improvements in disentanglement quality, cross-task generalization, and interpretability of the learned latent representations.

This work addresses the misalignment between disentangled representations in existing music generation models and their intended semantic meanings, which limits controllable generation. It presents the first multidimensional disentanglement evaluation framework for music audio, assessing representations along four axes—informativeness, equivariance, invariance, and disentanglement—through probing analyses and a range of unsupervised disentanglement strategies, including inductive biases, data augmentation, adversarial objectives, and staged training. The study systematically evaluates how well key attributes such as musical structure and timbre are captured in learned representations. Experimental results reveal significant discrepancies between the actual behavior of latent embeddings and their design intent, highlighting substantial challenges in achieving semantic consistency and offering new directions for improving controllable music generation mechanisms.