A5100433382- Underwater Acoustics Research
- Marine animal studies overview
- Acoustic Wave Phenomena Research
- Arctic and Antarctic ice dynamics
- Underwater Vehicles and Communication Systems
- Speech Recognition and Synthesis
- Metamaterials and Metasurfaces Applications
- Acoustic Wave Resonator Technologies
- Ferroelectric and Piezoelectric Materials
- Speech and Audio Processing
- Animal Vocal Communication and Behavior
- Music and Audio Processing
- Microwave Dielectric Ceramics Synthesis
- Geomechanics and Mining Engineering
- Marine and fisheries research
- Sugarcane Cultivation and Processing
- Semiconductor Lasers and Optical Devices
- Geoscience and Mining Technology
- Natural Language Processing Techniques
- Semiconductor Quantum Structures and Devices
- Soil Carbon and Nitrogen Dynamics
- Ichthyology and Marine Biology
- Ultrasonics and Acoustic Wave Propagation
- Bioenergy crop production and management
- Biomimetic flight and propulsion mechanisms
Xiamen University
2016-2025
Harbin Institute of Technology
2016-2025
Shenyang Institute of Engineering
2024
Massachusetts Institute of Technology
2013-2024
Wuhan National Laboratory for Optoelectronics
2024
Huazhong University of Science and Technology
2005-2024
China University of Petroleum, East China
2024
State Key Laboratory of Marine Environmental Science
2014-2024
Beth Israel Deaconess Medical Center
2024
Weatherford College
2024
We present SpecAugment, a simple data augmentation method for speech recognition. SpecAugment is applied directly to the feature inputs of neural network (i.e., filter bank coefficients). The policy consists warping features, masking blocks frequency channels, and time steps. apply on Listen, Attend Spell networks end-to-end recognition tasks. achieve state-of-the-art performance LibriSpeech 960h Swichboard 300h tasks, outperforming all prior work. On LibriSpeech, we 6.8% WER test-other...
We present a factorized hierarchical variational autoencoder, which learns disentangled and interpretable representations from sequential data without supervision. Specifically, we exploit the multi-scale nature of information in by formulating it explicitly within graphical model that imposes sequence-dependent priors sequence-independent to different sets latent variables. The is evaluated on two speech corpora demonstrate, qualitatively, its ability transform speakers or linguistic...
Domain mismatch between training and testing can lead to significant degradation in performance many machine learning scenarios. Unfortunately, this is not a rare situation for automatic speech recognition deployments real-world applications. Research on robust be regarded as trying overcome domain issue. In paper, we address the unsupervised adaptation problem recognition, where both source target are available, but word transcripts only available speech. We present novel augmentation-based...
Quasi-2D (Q-2D) perovskites are promising candidates to apply in light-emitting diodes (LEDs). However, delicate control on crystallization kinetics is needed suppress severe phase segregation. Here, the of Q-2D investigated via situ absorbance spectroscopy and for first time find multiphase distribution governed by arrangement, rather than diffusion, spacer cations at nucleation stage, which associate with its assembling ability determined molecular configuration. A "co-assembly" strategy...
Aberration layers (AL) often present significant energy transmission barriers in microwave engineering, electromagnetic waves, and medical ultrasound. However, achieving broadband ultrasonic focusing through aberration like the human skull using conventional materials such as metals elastomers has proven challenging. In this study, we introduce an inverse phase encoding method employing tunable soft metalens to penetrate heterogeneous layers. Through application of effective-medium theory,...
A bioinspired metagel achieves broadband tunable acoustic impedance matching.
Under thermal stress, perovskite materials suffer from volatile component loss or ion migration, etc., which is challenging for steady power output (SPO) of the resulting solar cells (PSCs) under practical operation conditions. Herein, we innovatively introduce silicon dioxide particles at perovskite/hole transport layer interface, simultaneously serve as heat dissipation material due to their higher conductivity and a surface passivator through coordination between undercoordinated lead...
Abstract Materials capable of varying their acoustic impedances to match those air, water, and solid materials over broadband frequencies will enable new applications in fields as diverse medical imaging, underwater sonar communication, marine biology. However, such tunability has not been achieved with conventional including metals, polymers, ceramics woods. Here, the invention metagel, a class designed hydrogel composites unprecedented tunable properties frequencies, is reported. The...
We present the Latent Sequence Decompositions (LSD) framework. LSD decomposes sequences with variable lengthed output units as a function of both input sequence and sequence. training algorithm which samples valid extensions an approximate decoding algorithm. experiment Wall Street Journal speech recognition task. Our model achieves 12.9% WER compared to character baseline 14.8% WER. When combined convolutional network on encoder, we achieve 9.6%
Nano Bi<sub>2</sub>WO<sub>6</sub> used as burning catalyst of NG/NC propellant increase rate, decrease pressure exponent and form specific high-pressure platform.
One-way propagation of acoustic waves has been achieved using superlattices or nonlinear diodes, but these structures are much larger than the wavelength. This makes them impractical for applications such as low-frequency underwater signal transmission. The authors show that unidirectional wave is possible with a device smaller Their design offers high rectification and broadband performance, does not require an external power source.
The uniqueness of soft materials such as hydrogels allows for great potential new robots and actuators. Conventional underwater are generally crafted rigid structures with water-mismatched acoustic impedance, severely limiting their operating capabilities. Therefore, flexible acoustically transparent show prospects applications in marine exploration, biomedical engineering, etc. Inspired by the high water content jellyfish, we report a hydraulically actuated bioinspired hydrogel jellyfish....
Sonar remains a major way to detect and discriminate underwater targets by interpreting the echoes. In this study, we used broadband dolphin clicks classify targets. The peak notch features of echo spectra were coded, echoes obtained using five-click trains, with number changing from 1 50. Codes containing target interpretation classified convolutional neural networks (CNNs). Compared single click, increasing 5, 10, 20, 50 in train would gradually improve classification rate 3%, 6.1%, 8.2%,...
The fluvial acoustic tomography (FAT) system relies on the arrival time of signal to calculate parameters region. traditional method uses matching filter peak position received after cross-correlation calculation within a certain as point, but this is difficult be effectively applied complex underwater environment, especially in case extremely low SNR. To solve problem, two-channel deep learning model (DCA-Net) proposed detect chromatographic signals. Firstly, an interactive module designed...
In the context of “dual carbon” goals, energy structures are rapidly shifting towards cleaner, low-carbon solutions. The clean and efficient electrode boiler, with its unique heat generation mechanism, is well aligned this trend. This review begins by outlining operating principles boilers, emphasizing their advantages in terms efficiency environmental sustainability. It then examines current status boiler applications within framework objectives, addressing key challenges technological...