Purpose Small-scale devices often provide too few microphones for effective array processing. This study investigates geometry-aware unsupervised loss functions for a neural network-based virtual microphone estimator (NN-VME), aiming to generate virtual microphone (VM) signals without ground-truth recordings at VM positions. Design/methodology/approach Known array geometry is used to construct training constraints. Alongside a previously proposed loss that maximizes linear correlation between estimated VM signals and recordings from other channels, a new loss based on the time difference of arrival (TDOA) of each source is introduced. The losses are compared on simulated multi-source LibriSpeech data, and the generated VM signals are evaluated with multiple beamformers. Findings The correlation-based method generates high-quality VM signals and improves beamforming performance on CHiME-4. The LibriSpeech experiments demonstrate the effectiveness of geometry-aware unsupervised NN-VME under multi-source conditions and show that the generated VM signals can support different beamforming methods. Research limitations/implications The framework assumes known array geometry, and correlation- and TDOA-based constraints may become less reliable under highly reverberant conditions. Future work should improve robustness in diverse real-world environments. Originality/value This study systematically compares geometry-aware unsupervised NN-VME losses across tasks, introduces a TDOA-based constraint for multi-source target speech extraction, and evaluates the applicability of estimated VM signals across beamformers.
Purpose This paper aims to investigate the efficiency and viability of online pair selection methods for preference score collection in Speech Quality Assessment (SQA). It addresses the challenge of high annotation costs in traditional pairwise comparisons and explores whether preference data collected in a sparse manner can successfully produce accurate rankings and support robust SQA model training. Design/methodology/approach We propose an improved aggregation method to derive system-level scores from raw preference data. To equalize uncertainty across system pairs, we introduce a novel bias-based pair selection method. Furthermore, we develop a pseudo-labeling strategy that utilizes unlabeled utterances to infer missing preferences, thereby augmenting the training data and enhancing SQA model performance. Findings Experiments demonstrate that the proposed bias-based pair selection method achieves the most robust performance. Additionally, the proposed pseudo-labeling approach effectively compensates for data sparsity, significantly improving the SQA model's performance by leveraging unlabeled datasets. Originality/value This paper fulfills an identified need for cost-effective SQA data collection. It bridges the gap between active pair selection and deep learning-based SQA model training by introducing a novel bias-based selection mechanism and a pseudo-labeling strategy that leverages system-level scores to predict missing preferences.
Perceptual image quality is a complex, multifaceted attribute influenced by factors such as sharpness, noise, contrast, and semantic content. Traditional image quality assessment methods measure overall quality but lack interpretable decompositions to isolate specific perceptual factors. In this work, we leverage the generative capacity of text-to-image diffusion models, which can be leveraged through textual inversion to decompose image quality into disentangled embeddings. By structuring these embeddings according to perceptual quality metrics such as Mean Opinion Score, our framework enables decomposing image quality while preserving semantic content. Moreover, this decomposition provides a principled way to investigate the underlying factors and essence of perceptual image quality, including the role of semantic content. To our knowledge, this is the first framework that decomposes image quality using the generative power of diffusion models.
This paper improves the stability and hardware-efficiency of zero-cost differentiable neural architecture search (NAS) for object detection, addressing two weaknesses of prior work: zero-cost rankings sensitive to SuperNet initialization and neglected edge-device deployment constraints. The authors propose IZeroDNAS, which integrates the NASWOT zero-cost metric into differentiable NAS. The authors introduce a SuperNet warm-up initialization, plus a depthwise-accelerated variant, to stabilize the rankings and a hardware-aware loss combining latency-aware and depth-aware terms built from a device look-up table. The authors evaluate on the ScaledYOLOv4 and YOLOv7 search spaces using Pascal VOC and COCO and deploy on Qualcomm RB5, Jetson TX2 and Jetson AGX Orin. Warm-up initialization reduces the mean average precision variance across random seeds from 5.687 to 0.0025 while raising accuracy. IZeroDNAS discovers shallower, wider models that are smaller and faster than manually designed detectors and the baseline ZeroDNAS, completing a search in under 5 h on a single RTX 2080, three times faster than comparable methods. The warm-up’s stabilizing effect is shown empirically rather than proven theoretically and a strictly fair comparison with proprietary methods such as YOLO-NAS is infeasible due to their undisclosed search spaces. Future work targets broader zero-cost proxies and more accurate device-performance predictors. IZeroDNAS is the first method to jointly tackle ranking instability and hardware-awareness in zero-cost differentiable NAS for object detection. Its warm-up and depthwise-accelerated variant offer a simple, modular, resource-efficient way to stabilize zero-cost metrics in large-scale SuperNets, benefiting practitioners deploying detectors on edge devices.
Major depressive disorder (MDD) is a globally prevalent mental illness. Functional connectivity (FC) analysis based on resting-state functional magnetic resonance imaging resting-state fMRI (rs-fMRI) provides new opportunities for neurobiological mechanism research and objective diagnosis of MDD. However, existing methods are prone to overfitting when handling high-dimensional FC data and predominantly rely on single-feature learning strategies, which limits their ability to fully leverage synergistic information between local connectivity patterns and global network features. To address these challenges, this study proposes a multi-task dual-branch denoising autoencoder (MT-DBDAE) framework for automatic MDD diagnosis. The framework uses a heterogeneous dual-pathway module (HDM) that integrates FC transformer and graph attention networks (GATv2) to achieve multi-scale feature extraction, applies hidden space mixup technology for data augmentation in latent space and finally jointly optimizes three objectives – classification, reconstruction and contrastive learning – through a multi-task learning strategy. Nested cross-validation experiments were conducted on the large-scale multi-center resting-state (REST)-meta-MDD data set, achieving favorable classification performance on both the CC200 and anatomical automatic labeling atlases. This effectively alleviates issues of high-dimensional FC data and medical data scarcity, providing an effective tool for objective MDD diagnosis.
