Real-time speech processing in complex acoustic environments requires a good balance between enhancement quality and computational efficiency. While hierarchical encoder–decoder architectures are widely adopted for their low complexity, the inherent frequency compression often leads to the loss of high-resolution details, such as harmonic structures and fine-grained sub-band temporal dynamics. In contrast, high-performance dual-path architectures that operate at full resolution are typically too computationally intensive for real-time deployment. To bridge this gap, we propose EiDPM, an efficient encoder–decoder architecture that integrates high-resolution dual-path modeling into a real-time pipeline. We introduce initial and residual dual-path modeling (Pre-DPM and Post-DPM) modules that operate at the original time–frequency resolution to preserve structural fidelity before down-sampling and restore fine-grained details during feature reconstruction. By utilizing a purely convolutional design, the framework leverages weight-sharing as a strong inductive bias to learn robust features with minimal computational overhead and memory footprint. Experimental results demonstrate that EiDPM achieves the highest speech signal quality (SIG) and overall quality (OVRL) scores among the compared restoration models on the ICASSP 2024 Speech Signal Improvement (SSI) Challenge benchmark. On VoiceBank+DEMAND, EiDPM outperforms the encoder–decoder and multi-scale baselines across all evaluation metrics and achieves performance comparable to the dual-path-based model, with slightly lower PESQ, CBAK, and COVL scores but substantially lower runtime. Specifically, it outperforms recent real-time methods in speech signal quality (SIG) and overall quality (OVRL). Furthermore, on the ICASSP SSI Challenge dataset, it maintains a real-time factor (RTF) of 0.27 when evaluated on a single-thread CPU, confirming its suitability for low-latency applications.
Deep neural networks have demonstrated remarkable performance in speech separation tasks. However, the high computational complexity and parameter count of current speech separation models limit their practical deployment in real-world scenarios. Achieving high separation performance while maintaining low model complexity remains a significant challenge. To address this challenge, we propose a novel speech separation method named MSHP-Net (Multi-Scale Hierarchical Perceptual Network). Built upon the powerful SepReformer architecture, MSHP-Net employs a multi-scale hierarchical perceptual encoder to efficiently extract audio features, and leverages efficient Transformer modules to collaboratively complement contextual dependencies. Experimental results on two public datasets, Libri2Mix and LRS2-2Mix, demonstrate that MSHP-Net achieves the best separation performance among the compared lightweight models on Libri2Mix: at a parameter count of 1.26M and a computational cost of 8.59G/s, it improves SI-SNRi by about 0.9dB over TIGER-small, and its larger configuration, MSHP-Net Large, attains the best accuracy among them on LRS2-2Mix.
Anomalous sound detection (ASD) identifies deviations from learned normal acoustic patterns and is increasingly used in intelligent acoustic sensing for industrial monitoring, medical diagnosis, energy systems, transportation safety, urban sensing, and smart infrastructure. Recent benchmarks and evaluation protocols have shifted ASD from normal-only machine-sound detection toward domain shift, domain generalization, first-shot deployment, and multi-sensor monitoring. This review provides a problem-oriented survey that connects dataset evolution, evaluation metrics, machine-sound properties, methodological paradigms, and deployment requirements. We summarize representative benchmarks and related acoustic anomaly datasets, review benchmark- and deployment-oriented evaluation criteria, and organize ASD methods into five paradigms: reconstruction-based normality learning, embedding structuring learning, statistical scoring and calibration, robust fusion and evidence integration, and first-shot and deployment adaptation. The taxonomy clarifies how anomaly scarcity, domain variation, score calibration, heterogeneous sensing, and deployment constraints shape method design. We further discuss open challenges related to limited anomalous data, score instability, incomplete metadata, edge computation, reproducibility, and interpretability. Overall, ASD is moving from benchmark-centered anomaly detection toward deployable acoustic intelligence, where robust representations, calibrated decisions, efficient implementation, multi-sensor evidence alignment, and actionable explanations are essential for reliable real-world monitoring.
Speech-based Alzheimer’s disease (AD) detection is commonly formulated as discrete classification, although diagnostic labels may also exhibit an ordered cross-sectional structure. This paper proposes Latent Manifold Rectification (LMR) for low-resource multimodal AD classification from speech. The framework combines a multimodal variational autoencoder with topology-aware latent rectification and trajectory-interpolated consistency regularization (TICR). For NCMMSC2021, HC, MCI, and AD are treated as supervised classes; an MCI-aware soft disease-axis constraint encourages flexible HC–MCI–AD ordering, and TICR operates along the adjacent HC–MCI and MCI–AD transitions. For ADReSS, where MCI labels are unavailable, the MCI-related term is disabled and TICR reduces to HC–AD interpolation. Under speaker-independent 10-fold cross-validation, the A+T TICR model achieves accuracies of 0.8487 ± 0.0233 on NCMMSC2021 and 0.8300 ± 0.0286 on ADReSS. Evaluation on the official ADReSS Challenge split yields 0.8400 ± 0.0126 over five seeds. The main classification experiments show higher mean performance for LMR than for the corresponding baselines, although the corrected paired comparisons do not establish statistically significant superiority. Separately, quantitative analysis of the NCMMSC2021 latent space reveals a strong but non-rigid HC–MCI–AD ordered component, with Spearman ρ=0.873 together with substantial off-axis variation. These findings distinguish the observed classification trends from the evidence for diagnostic organization in the learned representation and support LMR as a diagnosis-aware regularization framework for low-resource AD classification.
