Cabin Noise Separation Using an Improved Convolutional Time Domain Noise Separation Networks

Qunyi He, Haitao Wang, Xiangyang Zeng, Kean Chen, Ye Lei, Shuwei Ren

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Traditional acoustic signal separation techniques usually have high requirements for experimental conditions and economic costs. In this paper, an improved convolutional time domain noise separation networks is proposed for noise separation in the cabin environment based on the classical CONY-TasNet. Parallel dilated convolutions is designed in the separation layer to achieving relatively long time signal processing. Then, more envelope information can be obtained which is beneficial for enhancing the processing effect of cabin low-frequency noise and harmonic noise, and also reducing the loss of local noise information. Compared with other commonly used deep networks, the proposed networks has better separation results. It not only restores a more accurate spectrum structure, but also obtain smaller distortions.

Original languageEnglish
Title of host publication2023 6th International Conference on Information Communication and Signal Processing, ICICSP 2023
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages417-423
Number of pages7
ISBN (Electronic)9798350339994
DOIs
StatePublished - 2023
Event6th International Conference on Information Communication and Signal Processing, ICICSP 2023 - Xi'an, China
Duration: 23 Sep 202325 Sep 2023

Publication series

Name2023 6th International Conference on Information Communication and Signal Processing, ICICSP 2023

Conference

Conference6th International Conference on Information Communication and Signal Processing, ICICSP 2023
Country/TerritoryChina
CityXi'an
Period23/09/2325/09/23

Keywords

  • acoustic signal separation
  • cabin noise
  • deep networks
  • parallel dilated convolution

Fingerprint

Dive into the research topics of 'Cabin Noise Separation Using an Improved Convolutional Time Domain Noise Separation Networks'. Together they form a unique fingerprint.

Cite this