Unsupervised neural source separation for crowded environments based on deep Bayesian learning

2019 Fiscal Year Final Research Report

• Project/Area Number: 19K21553
• Project/Area Number (Other): 18H06490 (2018)
• Research Category: Grant-in-Aid for Research Activity Start-up
• Allocation Type: Multi-year Fund (2019); Single-year Grants (2018)
• Review Section: 1002:Human informatics, applied informatics and related fields
• Research Institution: National Institute of Advanced Industrial Science and Technology
• Principal Investigator: Bando Yoshiaki 国立研究開発法人産業技術総合研究所, 情報・人間工学領域, 研究員 (40828167)
• Project Period (FY): 2018-08-24 – 2020-03-31
• Keywords: 音環境認識 / 音源分離 / 深層ベイズ学習 / 償却変分推論

Outline of Final Research Achievements

This study aims to develop a framework that can train neural source separation by using multichannel audio signals without any supervision. We utilized the spatial information observed by a microphone array and trained the source separation based on the spatial likelihood of the separated signals. We first developed a deep Bayesian method that trains a neural source separation based on a complex Gaussian mixture model. This approach is then extended to an audio-visual source localization method that can deal with the diffuse noise and the unknown number of sound sources, which are problematic for recognizing the crowded real-world environments.

Free Research Field

統計的音響信号処理

Academic Significance and Societal Importance of the Research Achievements

従来の統計モデルの知見と、独立して研究されがちな近年の深層学習の知見を統合し、音源分離の教師なし学習の枠組みを実現した。これまで、深層学習に基づく音源分離には、個別の音源信号である正解データを大量に準備する必要があり、実世界の様々な音源を分離するには原理上限界があった。そこで、統計的信号処理で研究されてきたブラインド信号処理の知見を深層学習に導入することで、教師データを用いずとも音源分離を学習できる枠組みを実現した。