Search

Large-scale audio tagging datasets inevitably contain imperfect labels, such as clip-wise annotated (temporally weak) tags with no exact on- and offsets, due to a high manual labeling cost. This work proposes pseudo strong labels (PSL), a simple label augmentation framework that enhances the supervision quality for large-scale weakly supervised audio tagging. A machine annotator is first trained on a large weakly supervised dataset, which then provides finer supervision for a student model. Using PSL we achieve an mAP of 35.95 balanced train subset of Audioset using a MobileNetV2 back-end, significantly outperforming approaches without PSL. An analysis is provided which reveals that PSL mitigates missing labels. Lastly, we show that models trained with PSL are also superior at generalizing to the Freesound datasets (FSD) than their weakly trained counterparts., Accepted by ICASSP 2022

Keyword spotting (KWS) and speaker verification (SV) are two important tasks in speech applications. Research shows that the state-of-art KWS and SV models are trained independently using different datasets since they expect to learn distinctive acoustic features. However, humans can distinguish language content and the speaker identity simultaneously. Motivated by this, we believe it is important to explore a method that can effectively extract common features while decoupling task-specific features. Bearing this in mind, a two-branch deep network (KWS branch and SV branch) with the same network structure is developed and a novel decoupling feature learning method is proposed to push up the performance of KWS and SV simultaneously where speaker-invariant keyword representations and keyword-invariant speaker representations are expected respectively. Experiments are conducted on Google Speech Commands Dataset (GSCD). The results demonstrate that the orthogonality regularization helps the network to achieve SOTA EER of 1.31% and 1.87% on KWS and SV, respectively., Comment: Accepted at ICASSP 2022

The future regulation system puts forward higher requirements for the storage and management of large power data. According to the natural topological structure, dynamic change and intrinsic complex data characteristics of power grid model the knowledge map of power grid model is stored by introducing graph database Neo4j, which promotes the evolution of dispatching automatic analysis and decision- making algorithm to knowledge guidance. The key to build knowledge map is to extract the knowledge unit of power grid model accurately. By studying the data storage mode of Neo4j, this paper proposes 0 general method of building knowledge map of power grid model based on Python and storing it in graph database. The feasibility of the method is verified by practical cases, and the internal relationship of model data in knowledge map is explored by using Cypher query language.

A typical radial ungrounded distribution network model is set up in the 10kV true type distribution network fault simulation laboratory. Through a series of single-phase-to-ground fault tests, the zero sequence and three-phase voltage and current waveforms of traveling wave and power frequency are obtained. Wavelet multiresolution analysis is used to decompose the current and voltage traveling waves. The local modulus maxima of the wavelet transform can extract the polarity of the initial traveling wave. These waveforms prove that the current traveling wave magnitude is much more significant than power frequency current. Besides, the tests also prove that the initial zero sequence current and voltage waves in the faulty feeder are reverse in polarity.

Through the methods of physiological experiment in combination with acoustical experiment, the present research investigates the respiration and acoustic features of the Chinese learners during the task of self-narration. Twenty-four Chinese learners from Korea, Japan, America and Thailand are assigned to narrate in Mandarin Chinese. The spontaneous speech respiration which is greatly influenced by the physiological condition, cognitive processing, context and mood state tends to make the respiration curve to be complex. In this condition, compared with Chinese native speakers, different respiratory units have significant differences in inspiration/expiration duration, amplitude and slope in terms of respiratory parameters. The frequency of the abnormal pause is large in great extent, and the position of the abnormal pause is rather flexible, so its different categories correspond to the different respiration forms.

A new non-bionic method of lateral motion based on the quasi-passive research methods is presented which need not move center of mass (COM) to supporting foot in advance. The method can be achieved in the ordinary biped robot, which has a simple, controllable increasing quality and so on. Firstly, Process of lateral motion planning of the biped robot is introduced in detail. Then, trajectory of COM of the biped robot is analyzed by Lagrangian method and the feet and other parts are also planned in the different time of step. Finally, stability analysis and Optimized approach of the method has provided. The results of experiment show that the results of experiment in my paper show that the method is feasibility and validity.

A universal infrared remote-control signal decoder is designed and implemented within the multimedia teaching system on personal computer. The universal infrared remote-control signal decoder is used to transmit the decoded key code into personal computer by serial communication and control the multimedia teaching software on personal computer, such as the third party teaching software and self-developed teaching application, in order to make up an integrated multimedia teaching system.