Robust Acoustics and Speech Perception of Aerial Robot Under Ego Noise for Scene Understanding During Critical Emergency Response Missions
Abstract
Three key-techniques were investigated for achieving robust acoustics and speech perception of aerial robot for scene understanding during critical emergency response missions. For noise robust sound source localization, a noise robust desired sound direction estimation method is developed using LSTM based weighting function. The direction estimation experiments confirmed that the proposed method shows improved robustness under indoor surveillance noise environment characterized by presence of harmonic or nonstationary noise sources. For attaining signal enhancement under noisy environment, the GSC exploits spatial information and generates multi-channel enhanced signals on which the following DAE can act. As a result, the DAE can take advantage of the multi-channels by modeling the underlying relationship of the distortion with adjacent frequency bins in other frequencies and other channels. The evaluation results demonstrate that utilizing the results of the proposed GSC structure as an input to the DAE is effective in improving noise reduction and speech recognition performance. To improve acoustic event recognition performance and overcome the deficit of acoustic event resource, a novel DNN based transfer learning approach is developed. By utilizing the information transferred from the universal source domain, the proposed approach improved AEC accuracy in indoor surveillance experiments.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 04, 2018
- Accession Number
- AD1057694
Entities
People
- Hanseok Ko
- Minkyu Lee
- Sangwook Park
- Sungjae Lee
- Sungkyu Mun
Organizations
- Korea University