어텐션 기반 엔드투엔드 음성인식 시각화 분석

임성민,구자현,김회린 한국음성학회 2019 말소리와 음성과학 Vol.11 No.1

An end-to-end speech recognition model consisting of a single integrated neural network model was recently proposed. The end-to-end model does not need several training steps, and its structure is easy to understand. However, it is difficult to understand how the model recognizes speech internally. In this paper, we visualized and analyzed the attention-based end-to-end model to elucidate its internal mechanisms. We compared the acoustic model of the BLSTM-HMM hybrid model with the encoder of the end-to-end model, and visualized them using t-SNE to examine the difference between neural network layers. As a result, we were able to delineate the difference between the acoustic model and the end-to-end model encoder. Additionally, we analyzed the decoder of the end-to-end model from a language model perspective. Finally, we found that improving end-to-end model decoder is necessary to yield higher performance.

A Comprehensive Analysis of the End-to-End Delay for Wireless Multimedia Sensor Networks

Nasim Abbas,Fengqi Yu 대한전기학회 2018 Journal of Electrical Engineering & Technology Vol.13 No.6

Wireless multimedia sensor networks (WMSNs) require real-time quality-of-service (QoS) guarantees to be provided by the network. The end-to-end delay is very critical metric for QoS guarantees in WMSNs. In WMSNs, due to the transmission errors incurred over wireless channels, it is difficult to obtain reliable delivery of data in conjunction with low end-to-end delay. In order to improve the end-to-end delay performance, the system has to drop few packets during network congestion. In this article, our proposal is based on optimization of end-to end delay for WMSNs. We optimize end-to-end delay constraint by assuming that each packet is allowed fixed number of retransmissions. To optimize the end-to-end delay, first, we compute the performance measures of the system, such as end-to-end delay and reliability for different network topologies (e.g., linear topology, tree topology) and against different choices of system parameters (e.g., data rate, number of nodes, number of retransmissions). Second, we study the impact of the end-to-end delay and packet delivery ratio on indoor and outdoor environments in WMSNs. All scenarios are simulated with multiple runtimes by using network simulator-2 (NS-2) and results are evaluated and discussed.

A Comprehensive Analysis of the End-to-End Delay for Wireless Multimedia Sensor Networks

Abbas, Nasim,Yu, Fengqi The Korean Institute of Electrical Engineers 2018 Journal of Electrical Engineering & Technology Vol.13 No.6

Wireless multimedia sensor networks (WMSNs) require real-time quality-of-service (QoS) guarantees to be provided by the network. The end-to-end delay is very critical metric for QoS guarantees in WMSNs. In WMSNs, due to the transmission errors incurred over wireless channels, it is difficult to obtain reliable delivery of data in conjunction with low end-to-end delay. In order to improve the end-to-end delay performance, the system has to drop few packets during network congestion. In this article, our proposal is based on optimization of end-to end delay for WMSNs. We optimize end-to-end delay constraint by assuming that each packet is allowed fixed number of retransmissions. To optimize the end-to-end delay, first, we compute the performance measures of the system, such as end-to-end delay and reliability for different network topologies (e.g., linear topology, tree topology) and against different choices of system parameters (e.g., data rate, number of nodes, number of retransmissions). Second, we study the impact of the end-to-end delay and packet delivery ratio on indoor and outdoor environments in WMSNs. All scenarios are simulated with multiple run-times by using network simulator-2 (NS-2) and results are evaluated and discussed.

Differentiated CW Policy and Strict Priority Policy for Location-Independent End-to-End Delay in Multi-Hop Wireless Mesh Networks

BAE, Yun Han,KIM, Kyung Jae,PARK, Jin Soo,CHOI, Bong Dae The Institute of Electronics, Information and Comm 2010 IEICE TRANSACTIONS ON COMMUNICATIONS - Vol.93 No.7

<P>We investigate delay analysis of multi-hop wireless mesh network (WMN) where nodes have multi-channel and multiple transceivers to increase the network capacity. The functionality of the multi-channel and multiple transceivers allows the whole WMN to be decomposed into disjoint zones in such a way that <I>i)</I> nodes in a zone are within one-hop distance, and relay node and end nodes with different <I>CW<SUB>min</SUB></I>s contend to access the channel based on IEEE 802.11e EDCA, <I>ii)</I> different channels are assigned to neighbor zones to prevent the hidden node problem, <I>iii)</I> relay nodes can transmit and receive the packets simultaneously by multi-channel and multiple transceivers. With this decomposition of the network, we focus on the delay at a single zone and then the end-to-end delay can be obtained as the sum of zone-delays. In order to have the location-independent end-to-end delay to the gateway regardless of source nodes' locations, we propose two packet management schemes, called the <I>differentiated CW policy</I> and the <I>strict priority policy</I>, at each relay node where relay packets with longer hop count are buffered in higher priority queues according to their experienced hop count. For the <I>differentiated CW policy</I>, a relay node adopts the functionality of IEEE 802.11e EDCA where a higher priority queue has a shorter minimum contention window. We model a typical zone as a one-hop IEEE 802.11e EDCA network under non-saturation condition where priority queues have different packet arrival rates and different minimum contention window sizes. First, we find the PGF (probability generating function) of the HoL-delay of packets at priority queues in a zone. Second, by modeling each queue as <I>M/G/</I>1 queue with the HoL-delay as a service time, we obtain the packet delay (the sum of the queueing delay and the HoL-delay) of each priority queue in a zone. Third, the average end-to-end delay of packet generated at end node in each zone is obtained by summing up the packet delays at each zone. For the <I>strict priority policy</I>, we regard a relay node as a single queueing system with multiple priority queues where relay packets in priority queues are served in the order of strict priority. Relay node has smaller <I>CW<SUB>min</SUB></I> than end node has and relay node competes with end nodes in a zone. Using the PGF of HoL-delay of packet at relay node and end node, we obtain the packet delay in a zone. The average end-to-end delay to the gateway generated at end node in each zone is obtained. Finally, for both the differentiated CW policy and strict priority policy, by equating all end-to-end delays to be approximately equal, we find the minimum contention window sizes of each priority queue numerically by trial and error method so that end-to-end delays of packets are almost equal regardless of their source's location, respectively. Numerical results show that proposed two methods obtain almost same end-to-end delay of packets regardless of their generated locations and our analytical results are shown to be well matched with the simulation results.</P>

단측신경 접합시 신경외막의 제거가 신경섬유의 재생에 미치는 영향

박재환,김준식,박재우,박성근 大韓成形外科學會 1999 Archives of Plastic Surgery Vol.26 No.1

End-to-side neurorrhaphy had been firstly reported at the beginning of this century, but no further clinical reports have been released. In some cases of end-to-end neurorrhaphy, donor nerve incision or even sacrification is inevitably necessary. In doing end- to -side neurorrhaphy, has not been widely accepted because previous results were not satisfactory. The purpose of this study is to evaluate the applicability of end-to-side neurorrhaphy in rats. Twenty female Sprague-Dawley rats, weighing approximately 200-250 g(232.50±12.6g) were used. In 10 rats, their right and left peroneal nerves were cut and the distal end of peroneal nerves were end-to-side neurorrhaphied to the tibial nerve. On the right side, the epineurium was left intact and on the left side, it was removed. The proximal end of peroneal nerves were imbedded in the thigh muscle. After 6 months, eletrophysiologic tests, including nerve conduction velocity, the amplitude and the areas of action potential of the tibial cranial muscle were done and then all rats were sacrificed for histologic analysis. The other 10 offspring rats were raised and used for control group. The mean weight of the tibial cranial muscle was 1.74(±0.30)g on the right side, 1.54(±0.31)g on the left side and 1.97(±0.28)g in the control group, which was statistically not significant. The number of muscle fibers in x 100 magnification was 106.10(±4.28) on the right side, 109.67(±4.69) on the left side and 107.80(±4.64) in the control group, which was also not significant. The amplitude and width of muscle action potential were not statistically different among the right side, the left side and in the control group. However, the nerve conduction velocity was bilaterally delayed if compared to control group. In histologic analysis, the presence of epineurium did not impact the results and end-to-side neurorrhaphy were successfully regenerated in all rats. End-to-side neurorrhaphy seem to be a functionally and histologically tolerable alternative when end-to-end neurorrhaphy is not available.

End-to-End Quality of Service Constrained Routing and Admission Control for MPLS Networks

Oulai, Desire,Chamberland, Steven,Pierre, Samuel The Korea Institute of Information and Commucation 2009 Journal of communications and networks Vol.11 No.3

Multiprotocol label switching (MPLS) networks require dynamic flow admission control to guarantee end-to-end quality of service (QoS) for each Internet protocol (IP) traffic flow. In this paper, we propose to tackle the joint routing and admission control problem for the IP traffic flows in MPLS networks without rerouting already admitted flows. We propose two mathematical programming models for this problem. The first model includes end-to-end delay constraints and the second one, end-to-end packet loss constraints. These end-to-end QoS constraints are imposed not only for the new traffic flow, but also for all already admitted flows in the network. The objective function of both models is to minimize the end-to-end delay for the new flow. Numerical results show that considering end-to-end delay (or packet loss) constraints for all flows has a small impact on the flow blocking rate. Moreover, we reduces significantly the mean end-to-end delay (or the mean packet loss rate) and the proposed approach is able to make its decision within 250 msec.

백서에서 단-측면 신경문합술의 실험적 연구

이동근,김우경,이병일,김응구,동은상 大韓成形外科學會 1998 Archives of Plastic Surgery Vol.25 No.7

We investigate the efficacy of the end-to-side neurorrhaphy, and an alternative reconstructive method for the wide nerve gap comparing with the state of normal, denervated,and nerve graft. Thirty Sprague-Dawley rats were divided into 3 group of each 10 rats. In the right side of all groups, nerve branch to the gastrocnemius muscle was severed and the distal ending of the nerve was sutured was sutured to the lateral side of the tibial nerve after round excision of the epineurium. In the first group(normal control group), left side was untouched. In the second(denervated group), left side was sectioned at the distal 5mm from bifurcation of branch to gastrocnemius from sciatic nerve. In the third group(nerve graft group), left side was sectioned at two points, 5 mm and 15 mm distal to bifurcation of the branch to gastrocnemius nerve from sciatic nerve, and then cut ends were anastomosed again as a nerve graft. 5 months later, all animals were subject to electrophysiologic tests and histologic examination including the number of nerve fibers and the changes of muscle fibers. In the group I, the conduction velocity of the end-to-side neurorrhaphy side was delayed comparing to the normal control side(p < 0.05). But, there was no significant difference between end-to-end neurorrhaphy and nerve graft. In histologic finding, the amount of nerve fibers in the end-to-side neurorrhaphy was 50.8% of normal control(p <0.01), and 48.6% of nerve graft side(p <0.006). The histolgic findings of the gastrocnemius muscle in the stat of the end-to-side neurorrhaphy were similar to findings of the nerve graft side, and though slightly decreased size and deformed shape of the muscle fibers were noted, the architecture of the normal muscle configuration was sustained. In conclusion, these results demonstrated the end-to-side neurorrhaphy conducted the electrical stimuli and allowed the regeneration of axons from lateral surface of healthy main peripheral nerve. Therefore, end-to-side neurorrhaphy was thought to be and alternative method for the reconstruction of the wide nerve gap.

End-to-End Quality of Service Constrained Routing and Admission Control for MPLS Networks

Desire Oulai,Steven Chamberland,Samuel Pierre 한국통신학회 2009 Journal of communications and networks Vol.11 No.3

Multiprotocol label switching (MPLS) networks require dynamic flow admission control to guarantee end-to-end quality of service (QoS) for each Internet protocol (IP) traffic flow. In this paper, we propose to tackle the joint routing and admission control problem for the IP traffic flows in MPLS networks without rerouting already admitted flows. We propose two mathematical programming models for this problem. The first model includes end-to-end delay constraints and the second one, end-to-end packet loss constraints. These end-to-end QoS constraints are imposed not only for the new traffic flow, but also for all already admitted flows in the network. The objective function of both models is to minimize the end-to-end delay for the new flow. Numerical results show that considering end-to-end delay (or packet loss) constraints for all flows has a small impact on the flow blocking rate. Moreover, we reduces significantly the mean end-to-end delay (or the mean packet loss rate) and the proposed approach is able to make its decision within 250 msec.

요양보호사의 죽음에 대한 태도, 임종간호 인지 및 교육요구도

정은영,공정현,서지혜 대한통합의학회 2017 대한통합의학회지 Vol.5 No.4

Purpose : The current study objective was to determine the attitudes of nursing students to death, perception of end-of life care, and perceptions of educational training needs in relation to the provision of end-of life care, and to apply the results to the development of an efficient and effective education program in this regard. Method : Data collection was carried out between March and April 2017, and was achieved through the administration of a questionnaire to 124 caregivers working in a nursing home in C. A self-report instrument was used to measure nursing attitudes to death, perceptions of end-of-life care, and perceptions of educational training needs in relation to the provision of end-of-life care. The collected data were analyzed using SPSSⓇ Statistics for WindowsⓇ(version 21.0), and were calculated was whole numbers, percentages, mean±standard deviation, and Pearson’s correlation coefficient. Results : The average scores obtained were 2.85 out of a maximum of 4.00, 2.14 out of 4.00, and 2.42 out of 4.00, for attitudes to death, perceptions of end-of-life, and perceptions of educational training needs in relation to the provision of end-of-life care, respectively. A positive correlation was found between attitudes to death and perceptions of end-of-life care, and a negative correlation was found between perceptions of end-of-life care and educational training needs in relation to the provision of end-of-life care. Conclusion : The development and implementation of an educational program is necessary to ensure positive attitudes to death in nursing students and improve their perceptions about end-of-life care. Further research is also warranted to determine the impact of such a program in this regard.

종단형 음성인식에서의 인식 유닛에 따른 언어모델 연구

전형배,강병옥,정훈,오유리,이윤경,강점자,박전규 한국정보과학회 2021 정보과학회 컴퓨팅의 실제 논문지 Vol.27 No.6

In this paper, we investigate a recognition unit to learn an end-to-end speech recognition model. In end-to-end speech recognition, we can model the output unit as characters (or syllable in Korean), part-of-speech (POS), and words. We compare the character unit and wordpiece model (WPM) that can be trained from the data. End-to-end speech recognition can combine the score of an end-to-end model with the score of an external language model (LM). For LM fusion decoding, we train recurrent neural network (RNN) language models with the appropriate language model level (character/POS/word/WPM). An end-to-end speech recognition model with a character unit is converged fast because of a smaller output unit number. End-to-end speech recognition with the character unit has lower accuracy than end-to-end speech recognition with the wordpiece unit. With the language model in decoding, two models have comparable performance and the end-to-end model with wordpiece unit has an advantage in terms of the decoding speed. 본 논문에서는 종단형 음성인식에서의 인식 유닛 선정과 인식 성능과의 관계를 영어 및 한국어 음성인식 환경에서 살펴보았다. 종단형 음성인식의 인식 유닛은 글자(character), 형태소, 단어 등으로 정의하여 모델 학습을 할 수 있다. 본 논문에서는 이 중 기본이 되는 글자(character) 유닛과 데이터로부터 정의 가능한 wordpiece model(WPM)을 비교 실험하였다. 종단형 음성인식 방법은 디코딩 과정에서 외부 언어모델과 fusion을 통한 디코딩이 가능하다. 글자 유닛과 wordpiece 유닛에 대해 적절한 언어모델을 학습 후 언어모델 fusion 기반 음성인식을 수행하였다. 글자 유닛을 적용하는 종단형 음성인식의 경우 종단형 모델의 출력 유닛 개수가 적어 빠르게 모델이 수렴하는 특징이 있다. 영어 음성인식 환경에서는 글자 유닛을 적용한 종단형 모델 성능이 wordpiece 유닛을 적용한 종단형 모델 보다 낮게 관측되었다. 적절한 언어모델을 적용하여 디코딩을 수행하는 경우 양쪽 모두 유사한 최종 성능을 얻었으나, 언어모델의 유닛 개수 차이로 인하여 wordpiece 모델을 출력 유닛으로 하는 종단형 음성인식 모델이 인식 속도 측면에서 상대적으로 좋은 성능을 보였다.