An end-to-end (e2e) text-to-speech (TTS) system is a deep architecture that learns to associate a text string with acoustic speech patterns from a dataset. It is expected that all aspects associated with speech production, such as phone duration, speaker characteristics, and intonation among other things are captured in the trained TTS model. Human speech is complex, involving smooth transitions between articulatory configurations (ACs). Due to anatomical constraints, some ACs are challenging to mimic or transition between. In this paper, we pose the question “Do e2e-TTS systems benefit by the constraints imposed by human anatomy?”. To answer this, we build systems using reverse text and reverse speech. Experiments demonstrate that e2e TTS systems are purely data-driven. Interestingly, the generated speech by reverse systems exhibits better fidelity, better perceptual intelligibility, and better naturalness.
Prohibited items detection is crucial for maintaining security in public areas like airports and train stations. Despite progress in target detection and semantic segmentation algorithms that aid in identifying contraband, X-ray imaging presents unique difficulties due to the severe distortion of pixel information caused by object stacking and density variations. Many existing detection methods fail to account for the complex interplay between foreground and background pixels, resulting in subpar performance. We introduce XMamba, a specialized architecture derived from the VMamba model, designed specifically for detecting prohibited items in X-ray images. XMamba incorporates an Equalization Perception Module (EPM) and a Foreground Channel Decoupling (FCD) method to tackle the challenges of weak features, internal pseudo-contours, and perspective interferences. By leveraging state-space modeling (SSM) to reduce self-attention complexity, XMamba effectively captures global information and minimizes distortion caused by these challenges. Our approach strikes a balance between parameter count, GFLOPs, and accuracy, outperforming both CNN-based and transformer-based detectors. Experimental results demonstrate that our proposed XMamba detector, when combined with cascaded Mask R-CNN, comprehensively surpasses all state-of-the-art methods.
With the progression of an aging society, the number of patients with gait disorders is increasing. Representative conditions causing such disorders include lumbar spinal canal stenosis and cervical spondylotic myelopathy. This study aims to estimate these diseases from gait videos recorded from a lateral view of the patient. However, capturing gait videos typically requires a large space, and when recorded in a limited environment, the videos may suffer from occlusion and distortion, potentially degrading estimation accuracy. To address this issue, we propose a robust disease estimation method using Deep Mutual Learning (DML), where multiple models learn while sharing knowledge with each other. In the proposed framework, two deep learning models with shared parameters are used: one receives complete gait data and the other receives incomplete gait data. The models are trained to extract similar features from both types of input, enabling reliable disease estimation even from degraded videos. We demonstrate the effectiveness of the proposed method through evaluation experiments using gait videos of patients with gait disorders and healthy individuals.
Conventional aging simulations require the collection of facial images by age group; thus, a lack of training data poses an issue. As a result, generative AI that learns from big data has attracted attention. However, generative AI has the characteristic of generating plausible results but not necessarily accurate and has the drawback of lacking objectivity. This research aims to perform an objective aging simulation while leveraging the diverse facial expression generation capabilities of generative AI trained on big data to compensate for the lack of training data. When compared with aging simulations that utilizes prompt instructions with generative AI, the results tended to be uniform aging simulations for all individuals. However, the proposed method results in an aging simulation that reflects individual characteristics, and it was confirmed that an objective simulation can be realized while taking advantage of generative AI's ability to generate plausible images.
In the context of smart agriculture, effective monitoring of crop growth requires a 3D point cloud compression method that preserves detailed features of the primary plant structures while suppressing less relevant background information. However, achieving this balance is challenging due to the uniform treatment of spatial regions in conventional compression schemes. To address this need, we propose a resource-efficient importance-aware adaptive compression framework for point clouds. Specifically, after performing semantic segmentation on the raw point cloud, we quantify the semantic importance of each segmented region and map the importance scores to adaptive compression ratios. These ratios are then embedded into the latent representation during encoding and parsed during decoding to guide the reconstruction precision. This enables a more efficient allocation of bits to semantically critical regions. We validate the effectiveness of the proposed method on both our self-collected crop dataset and the publicly available Pheno4D dataset, demonstrating its feasibility and improved performance in balancing compression rate and reconstruction accuracy.
