Search

A Multilingual Keyword Spotting (KWS) system detects spokenkeywords over multiple locales. Conventional monolingual KWSapproaches do not scale well to multilingual scenarios because ofhigh development/maintenance costs and lack of resource sharing.To overcome this limit, we propose two locale-conditioned universalmodels with locale feature concatenation and feature-wise linearmodulation (FiLM). We compare these models with two baselinemethods: locale-specific monolingual KWS, and a single universalmodel trained over all data. Experiments over 10 localized languagedatasets show that locale-conditioned models substantially improveaccuracy over baseline methods across all locales in different noiseconditions.FiLMperformed the best, improving on average FRRby 61% (relative) compared to monolingual KWS models of similarsizes., Comment: Accepted for ICASSP 2023

The behavior of bubbles traveling in the proximity of a tilted wall is studied experimentally to understand the fundamental sliding motion of bubbles inside turbulent boundary layers along an inclined wall. The qualitative visualization of sliding bubbles confirms the contribution of bubble buoyancy on the sliding motion for negative and positive inclinations of the channel. An opto-acoustic combined measurement technique is adopted to explore the sliding motion. Liquid velocity profiles in the bubbly flow and the distance between the wall and bottom of the bubble are obtained using the ultrasound pulsed Doppler method, while the bubble diameters and velocities are obtained from particle-tracking type image processing. The combined measurements reveal that the velocity of bubbles decreases under the negative slope condition and increases under the positive slope condition due to opposite buoyancy effects. In addition, the distance between the wall and bottom of the bubble increases with an increase in negative inclination. The lift coefficient is derived from the measured variables using a force-balance equation among the buoyancy, lift, and surface tension. Finally, we propose modeling equations for the lift coefficient expressed in terms of the Reynolds, Weber, and Bond numbers, which apply to the bubbles inside boundary layers.

A velocity vector profile technique based on an ultrasound pulsed Doppler method can enrich the information on the flow field. However, it has shown low availability because a new design of special transducers is required for each measurement case. This study proposes a new method of profiling the velocity vectors using conventional ultrasound transducers that are widely supplied to ultrasound velocity profile users. We constructed a configuration of transducers to minimize the uncertainty of the detection points at the receivers, and a measurable distance was theoretically determined by the configuration. Two feasibility tests were carried out. One was a test for the assessment of the measurable distance, which agreed well with the theoretical distance. The other was the evaluation of the measurement of 2D velocity vectors by the new method, and it was performed in a towing tank facility without the velocity fluctuation. From the evaluation, it was confirmed that the measured vectors showed good agreement with the reference values, and their accuracy and precision were competitive compared to previous studies. In order to demonstrate this, the developed method was applied to two unsteady flows. The results clarified that the proposed method guarantees high availability and accuracy for the velocity vector profiles., source:https://doi.org/10.1088/1361-6501/ac37ec

Owing to different flow conditions-for example, Poiseuille and Couette flow-one could expect different deformations of large-sized bubbles; however, bubble dynamics has been mostly investigated in channel flow. Consequently, an intermediate flow condition shared by both the channel and ship surface is needed for large-sized bubbles, although it can be difficult to simultaneously generate turbulent Couette flow in the channel and measure the shear stress on the ship's surface experimentally. In this study, large-sized bubbles in turbulent Couette flow were investigated numerically to determine their common characteristics under such flow conditions. The interlsoFoam solver from OpenFOAM-which can directly capture the interface via the geometric volume of fluid method-was used to conduct the simulations of the gas-liquid interface problem. The turbulent Couette flow was driven by top wall velocity condition with an initial perturbation, and three different bubble sizes with Weber numbers in the range of 200-300 were chosen to determine the characteristics of large-sized bubbles. By monitoring the results according to bubble size, we could determine bubble characteristics that were distinguishable from those in turbulent Poiseuille flow. Consequently, bubble deformation was dominated by the velocity gradient and shear rate, which was greater than that during single-phase flow from the liquid-film region. These results allowed us to generalize the deformation mechanism of large-sized bubbles in turbulent Couette flow into five categories-namely, the initial shape, deformation on the front side, change of the center of gravity, pinch-off/breaking of the ligament, and deformation to a stable shape.

Jun Soo Park,1 Min Seop Kim,1 Min Yeong Joung,1 Hyun Jin Park,1 Myoung-Jin Ho,1 Jun Hyuk Choi,1 Jae Hee Seo,1 Woo Heon Song,1 Young Wook Choi,2 Sangkil Lee,3 Yong Seok Choi,1 Myung Joo Kang1 1College of Pharmacy, Dankook University, Cheonan, Republic of Korea; 2College of Pharmacy, Chung-Ang University, Seoul, Republic of Korea; 3College of Pharmacy, Keimyung University, Daegu, Republic of KoreaCorrespondence: Myung Joo Kang, College of Pharmacy, Dankook University, Dongnam-gu, Cheonan, 31116, Republic of Korea, Tel +82 41 550 1446, Fax +82 41 550 7899, Email kangmj@dankook.ac.krBackground: Montelukast (MTK), a representative leukotriene receptor antagonist, is currently being investigated as a potential candidate for treating Alzheimer’s disease. For potent and effective dosing in elderly patients, a parenteral prolonged delivery system is favored, with improved medication adherence with reduced dosage frequency.Purpose: This study aimed to design a nanocrystalline suspension (NS)-based MTK prolonged delivery system and evaluate its pharmacokinetics profile and local tolerability following subcutaneous administration.Methods: To decelerate the dissolution rate, the amorphous MTK raw material was transformed into a crystalline state using a solvent-mediated transformation method and subsequently formulated into NS using a bead-milling technique. The MTK NSs were characterized by morphology, particle size, crystallinity, and in vitro dissolution profiles. The pharmacokinetic profile and local tolerability at the injection site following subcutaneous injection of MTK suspension were evaluated in rats.Results: Microscopic and physical characterization revealed that the amorphous MTK powder was lucratively transformed into a crystalline form in acidic media (pH 4). MTK crystalline suspensions with different diameters (200 nm, 500 nm, and 3 μm) were uniformly prepared using bead-milling technology, employing polysorbate 80 as suspending agent. Prepared c

We trained a keyword spotting model using federated learning on real user devices and observed significant improvements when the model was deployed for inference on phones. To compensate for data domains that are missing from on-device training caches, we employed joint federated-centralized training. And to learn in the absence of curated labels on-device, we formulated a confidence filtering strategy based on user-feedback signals for federated distillation. These techniques created models that significantly improved quality metrics in offline evaluations and user-experience metrics in live A/B experiments., Comment: Accepted to Interspeech 2022

Eun A Kim, Jun Soo Park, Min Seop Kim, Min Young Jeong, Hyun Jin Park, Jun Hyuk Choi, Jae Hee Seo, Yong Seok Choi, Myung Joo Kang College of Pharmacy, Dankook University, Cheonan, Chungnam, 330-714, KoreaCorrespondence: Myung Joo KangCollege of Pharmacy, Dankook University, 119 Dandae-ro, Dongnam-gu, Cheonan, 330-714, KoreaTel +82 41 550 1446Fax +82 41 550 7899. Email kangmj@dankook.ac.krBackground: The titrated extract of Centella asiatica (CA) has received much attention as a cosmeceutical ingredient owing to its anti-wrinkle effect. However, due to the low solubility and high molecular weight of pharmacologically active constituents, including asiatic acid (AA), madecassic acid (MA), and asiaticoside (AS), it is challenging to fabricate high-payload topical preparations of CA with satisfactory skin absorption profiles.Purpose: This study aimed to design a high-payload topical preparation of CA using nanocrystallization technique and to evaluate its skin absorption profile and local tolerability.Methods: High-payload nanocrystal suspensions (NSs) were prepared using lab-scale bead-milling technology, by adjusting the type and amount of suspending agent, CA content, type of vehicle, and milling speed. CA-loaded NSs were characterized in terms of morphology, particle size, crystallinity, and in vitro dissolution pattern. Skin absorption of CA nanocrystals was evaluated using a vertical Franz diffusion cell mounted with porcine skin. In vivo skin irritation following topical application of high-payload NS was assessed in normal rats.Results: The optimized NS system, composed of 10% (w/v) CA, 0.5% polyvinylpyrrolidone (PVP) K30 as steric stabilizer, and 89.5% of distilled water, was characterized as follows: spherical or elliptical in shape, 200 nm in size, with low crystallinity. The in vitro dissolution of AA or MA from NSs was markedly faster compared to raw material, under sink condition. Penetration of AA, MA, and AS in the porcine skin was markedly elevated usin

Rainbow particle tracking velocimetry (PTV) is a PTV method that enables three-dimensional (3D) three-component flow measurement using a single camera. Despite the advantage of its simple setup, the accuracy of the particle depth is restricted due to false color caused by image sensor arrays, such as Bayer arrangement. Since the false color occurs near sharp edges in the color gradient of in-focus individual particle images, we here introduced a defocusing technique to rainbow PTV to remove these false colors. Defocusing led to moon-shaped distorted particle images, which we applied an adaptive mask correlation technique to detect. Multi-cycle rainbow illumination was realized as an additional improvement on the defocusing technique. In particular, individual particle coordinates were obtained by a combination of the color and constitution of pixels. This dramatically increased the depth resolution of the 3D particle tracking. The feasibility of the proposed method was demonstrated by a flow driven by rotating impellers and a wake behind a twisted Savonius turbine. By the demonstration, it is confirmed that the twisted turbine suppresses the loss of kinetic energy by shedding streamwise vortices in the wake.

We propose self-training with noisy student-teacher approach for streaming keyword spotting, that can utilize large-scale unlabeled data and aggressive data augmentation. The proposed method applies aggressive data augmentation (spectral augmentation) on the input of both student and teacher and utilize unlabeled data at scale, which significantly boosts the accuracy of student against challenging conditions. Such aggressive augmentation usually degrades model performance when used with supervised training with hard-labeled data. Experiments show that aggressive spec augmentation on baseline supervised training method degrades accuracy, while the proposed self-training with noisy student-teacher training improves accuracy of some difficult-conditioned test sets by as much as 60%.

Although the practical use of bubbly drag reduction for marine vessels has begun to spread, it is unclear how bubbles migrate along spatially developing turbulent boundary layers. We measured the spatiotemporal distribution of the local void fraction beneath a 4-m-long fully transparent flat-bottom model ship towed in a 100-m-long water tank with a ship speed of up to 3.00 m/s. Bubbles were injected with both constant and periodically fluctuated air flow rates into the turbulent boundary layers. With constant flow rates, 2-5-mm bubbles caused spontaneous void waves, i.e., quasiperiodic fluctuation of the local void fraction, standing at 4-8 Hz. When repetitive bubble injection, which formed artificial void waves downstream, was applied, void waves propagated at 47%-53% of the ship speed. During the propagation, we found the apparent diffusion coefficient of the void wave taking values down to 10(-4) m(2)/s and even negative values when the repetition frequency was set lower than the spontaneous frequency of void waves. This indicates that the void waves persist for several tens of meter in application to ship hulls.

Following the success of bubbly drag reduction for marine vessels by hydrofoil bubble generators (Kumagai et al., 2015), we conducted two kinds of experiments to specify and improve its working principle. Firstly, PIV measurement was carried out to elucidate the flow structure around a hydrofoil running beneath a free surface. Froude number dependence of the anti-stall characteristics and cancellation of vertical pressure gradient above the hydrofoil were confirmed, which entrain the atmospheric air into water flow. Secondly, we mounted the hydrofoil to a model ship that runs in a 100-m-towing facility and found three different two-phase flow patterns in realizing bubble generation. Volume flow rate of generated bubbles increased with the ship speed to the power of greater than two, thus wall-occupation fraction of bubbles also increases with the ship speed. Furthermore, hydrofoil drag during bubble generation was measured and used to estimate net power-saving, having proved advantage of the hydrofoil bubble generator for long ships at high speed conditions.

Following the success of bubbly drag reduction for marine vessels by hydrofoil bubble generators (Kumagai et al., 2015), we conducted two kinds of experiments to specify and improve its working principle. Firstly, PIV measurement was carried out to elucidate the flow structure around a hydrofoil running beneath a free surface. Froude number dependence of the anti-stall characteristics and cancellation of vertical pressure gradient above the hydrofoil were confirmed, which entrain the atmospheric air into water flow. Secondly, we mounted the hydrofoil to a model ship that runs in a 100-m-towing facility and found three different two-phase flow patterns in realizing bubble generation. Volume flow rate of generated bubbles increased with the ship speed to the power of greater than two, thus wall-occupation fraction of bubbles also increases with the ship speed. Furthermore, hydrofoil drag during bubble generation was measured and used to estimate net power-saving, having proved advantage of the hydrofoil bubble generator for long ships at high speed conditions.

We propose smoothed max pooling loss and its application to keyword spotting systems. The proposed approach jointly trains an encoder (to detect keyword parts) and a decoder (to detect whole keyword) in a semi-supervised manner. The proposed new loss function allows training a model to detect parts and whole of a keyword, without strictly depending on frame-level labeling from LVCSR (Large vocabulary continuous speech recognition), making further optimization possible. The proposed system outperforms the baseline keyword spotting model in [1] due to increased optimizability. Further, it can be more easily adapted for on-device learning applications due to reduced dependency on LVCSR., Comment: Accepted in International Conference on Acoustics, Speech, and Signal Processing 2020

Three-dimensional distribution of microbubbles dispersed in a turbulent boundary layer spatially developing along a flat plate is investigated experimentally. The aim of study is to find out visually how microbubbles interact with turbulent eddies to achieve frictional drag reduction on a wall. Setting conditions of the experiments are stream Reynolds number up to 2.3 x 10(4), microbubble volume fraction inside boundary layer less than 10(-5), and microbubble Stokes number smaller than 10(-3). Stream-wise vortices in the boundary layer were visualized by flake optics, of which spacing expands with injection of microbubbles. 3-D microbubble distributions were visualized by a color-coded volumetric illumination. The result showed preferentially accumulated microbubble clouds to low-speed streaks close to viscous sublayer and to hair-pin vortices in the buffer layer, leading to reduced diffusivity of microbubbles relative to the turbulent momentum diffusivity.

We established a vortex detection method using instantaneous temperature fields that were visualized using thermochromic liquid crystals (TLCs) to investigate behaviors of vortical structures in a rotating Rayleigh-Benard convection. Experimental testing was performed at a fixed Rayleigh number in water containing encapsulated TLCs. Vortices were recognized as undulations that appear in the horizontal temperature fields, thus making vortex detection with high spatial resolution possible, and this enabled quantitative investigation of the dynamics of vortical structures. Standard template matching was used to detect individual vortices on visualized temperature fields, and two-dimensional curved surface fitting was adopted to remove erroneous detections and to evaluate shapes of local temperature fields corresponding to vortical structures. Additionally, vortex tracking clearly showed geometric advection pattern of vortical structures.

The combined use of ultrasound pulse-echo intensity and Doppler shift frequency is examined as a means to measure strong unsteady three-phase pipe flows of a gas and two liquids. With air, oil, and water as components of the fluid media, particular attention is given to analyze ultrasound responses at the air-oil and oil-water interfaces. Reciprocating slugging is generated inside a 55-mm-diameter circular pipe, of which edges oscillate vertically at a controlled frequency. We use an ultrasound velocity profiler to obtain the 1-D cross-sectional distributions of the instantaneous flow velocity at the sampling rate of 60 Hz. All the measurements are realized by a single ultrasound transducer located outside the pipe. Measurement accuracy is validated using a high-speed camera coupled with particle image velocimetry that is synchronized with the profiler. The results demonstrate that the proposed technique works properly in sensing both interfaces as well as in-phase flow velocity distributions. In addition, multiphase volume flow rates for the constituents are obtained by velocity profile integration assuming vertical phase stratification in an approximation.

A novel visualization technique was proposed to extract the three-dimensional vortex structure of a turbulent puff, which is a local turbulence event that is observed in pipe flows at relatively low Reynolds numbers. The technique is based on multi-color illumination of microscopic flakes that are suspended in the flow, which makes structural visualization more informative than conventional monochrome approaches. A special optical arrangement of two laser sheets, colored green and blue, was established for the circular pipe. Based on an image analysis sequence, the internal structure of the puff is reconstructed as a cross-sectional temporal 3D image consisting of voxels with unicolor degrees between green and blue, where an individual single vortex is extracted as a pair of two-color stripes. This allows quantification of the azimuthal wavenumber of the vortical structure that characterizes the puff. The wavenumber results agreed well with the results of previous studies, thus supporting the applicability of the proposed visualization technique. .

The injection of bubbles into a turbulent boundary layer can reduce the skin friction of a wall. Conventionally, the drag reduction rate is evaluated using time-averaged quantities of the mean gas flow rate or mean void fraction. Actually, as bubbles are subject to strong shear stresses near the wall, void waves and local bubble clusters appear. For pipe and channel flows, such wavelike behavior of the dispersed phase has been investigated intensely as an internal two-phase flow problem. We investigate how this wavy structure forms within the boundary layer as an external spatially developing two-phase flow along a horizontal flat plate. We describe how our model ship is designed to meet that purpose and report bubble-traveling behavior that accompanies unexpectedly strong wavy oscillations in the streamwise direction. A theoretical explanation based on a simplified two-fluid model is given to support this experimental fact, which suggests that void waves naturally stand out when drag reduction is enhanced through the local spatial gradient of the void fraction.

Capturing dependencies in images in an unsupervised manner is important for many image-processing applications and for understanding the structure of natural image signals. Data generative linear models such as principal component analysis (PCA) and independent component analysis (ICA) have shown to capture low level features such as oriented edges in images. However, those models capture only linear dependency and therefore its modeling capability is limited. We propose a new method for capturing nonlinear dependencies in images of natural scenes. This method is an extension of the linear ICA method and builds on a hierarchical representation. The model makes use of lower level linear ICA representation and a subsequent mixture of Laplacian distribution for learning the nonlinear dependencies in an image. The model parameters are learned via the expectation maximization (EM) algorithm and it can accurately capture variance correlation and other high order structures in a simple and consistent manner. We visualize the learned variance correlation structure and demonstrate applications to automatic image segmentation and image denoising. (c) 2006 Elsevier B.V. All rights reserved.