Generative Adversarial Network with Guided Generator for Non-stationary Noise Cancelation

Kyung Hyun Lim, Jin Young Kim, Sung Bae Cho

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Noise comes from a variety of sources in real world, which makes a lot of non-stationary noises, and it is difficult to find target speech from noisy auditory signals. Recently, adversarial learning models get attention for its high performance in the field of noise control, but it has limitation to depend on the one-to-one mapping between the noisy and the target signals, and unstable training process due to the various distributions of noise. In this paper, we propose a novel deep learning model to learn the noise and target speech distributions at the same time for improving the performance of noise cancellation. It is composed of two generators to stabilize the training process and two discriminators to optimize the distributions of noise and target speech, respectively. It helps to compress the distribution over the latent space, because two distributions from the same source are used simultaneously during adversarial learning. For the stable learning, one generator is pre-trained with minimum sample and guides the other generator, so that it can prevent mode collapsing problem by using prior knowledge. Experiments with the noise speech dataset composed of 30 speakers and 90 types of noise are conducted with scale-invariant source-to-noise ratio (SI-SNR) metric. The proposed model shows the enhanced performance of 7.36, which is 2.13 times better than the state-of-the-art model. Additional experiment on −10, −5, 0, 5, and 10 dB of the noise confirms the robustness of the proposed model.

Original languageEnglish
Title of host publicationHybrid Artificial Intelligent Systems - 15th International Conference, HAIS 2020, Proceedings
EditorsEnrique Antonio de la Cal, José Ramón Villar Flecha, Héctor Quintián, Emilio Corchado
PublisherSpringer Science and Business Media Deutschland GmbH
Pages3-12
Number of pages10
ISBN (Print)9783030617042
DOIs
Publication statusPublished - 2020
Event15th International Conference on Hybrid Artificial Intelligent Systems, HAIS 2020 - Gijón, Spain
Duration: 2020 Nov 112020 Nov 13

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12344 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference15th International Conference on Hybrid Artificial Intelligent Systems, HAIS 2020
CountrySpain
CityGijón
Period20/11/1120/11/13

Bibliographical note

Funding Information:
This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2020-0-01361, Artificial Intelligence Graduate School Program (Yonsei University)) and grant funded by 2019 IT promotion fund (Development of AI based Precision Medicine Emergency System) of the Korean government (MSIT).

Funding Information:
Acknowledgement. This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2020-0-01361, Artificial Intelligence Graduate School Program (Yonsei University)) and grant funded by 2019 IT promotion fund (Development of AI based Precision Medicine Emergency System) of the Korean government (MSIT).

Publisher Copyright:
© 2020, Springer Nature Switzerland AG.

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of 'Generative Adversarial Network with Guided Generator for Non-stationary Noise Cancelation'. Together they form a unique fingerprint.

Cite this