Cross-language transfer speech recognition using deep learning

Zhao, Yue; Xu, Yan; Sun, Mei J.; Xu, Xing; Wang, Hui; Yang, Guo; Ji, Qiang

doi:10.1109/icca.2014.6871132

Cited by 4 publications

(1 citation statement)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…One way to train a CNN for speech recognition is through the use of spectrograms as input images, (Li et al , 2013). Various techniques have been investigated to the end of building robust speech recognizers and/or to cater for multiple languages (Seki et al , 2017; Wu et al , 2017; Kundu et al , 2016; Chen and Mak, 2015; Zhao et al , 2014); however, in this context, a simple speech recognition was required, with the ability to categorize audio inputs as one of seven possibilities, namely, “Yes”, “No”, “Okay”, “Don’t”, “Wait”, “Stop” and Negative. The first six categories included two positive answers and four negative answers, given that the monitoring system shall ask predefined yes/no questions to the operator when in need of further clarification, and the negative category consisted of examples of background noise.…”

Section: Building the Individual Systemsmentioning

confidence: 99%

A neural network based monitoring system for safety in shared work-space human-robot collaboration

Rajnathsing

2018

View full text Add to dashboard Cite

Purpose Human–robot collaboration (HRC) is on the rise in a bid for improved flexibility in production cells. In the context of overlapping workspace between a human operator and an industrial robot, the major cause for concern rests on the safety of the former. Design/methodology/approach In light of recent advances and trends, this paper proposes to implement a monitoring system for the shared workspace HRC, which supplements the robot, to locate the human operator and to ensure that at all times a minimum safe distance is respected by the robot with respect to its human partner. The monitoring system consists of four neural networks, namely, an object detector, two neural networks responsible for assessing the detections and a simple, custom speech recognizer. Findings It was observed that with due consideration of the production cell, it is possible to create excellent data sets which result in promising performances of the neural networks. Each neural network can be further improved by using its mistakes as examples thrown back in the data set. Thus, the whole monitoring system can achieve a reliable performance. Practical implications Success of the proposed framework may lead to any industrial robot being suitable for use in HRC. Originality/value This paper proposes a system comprising neural networks in most part, and it looks at a digital representation of the workspace from a different angle. The exclusive use of neural networks is seen as an attempt to propose a system which can be relatively easily deployed in industrial settings as neural networks can be fine-tuned for adjustments.

show abstract