| Presentation | 2018-11-21 Improvement of speech distortion caused by multi-layered neural network-based beamformer with alternative training Kouhei Eguchi, Mitsunori Mizumachi, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | A neural network-based beamformer can simultaneously optimize its beam-pattern at the desired and undesired directions so that it enables noise reduction more efficiently compared with a conventional delay-and-sum beamformer with the linear signal processing. In general, the neural network-based beamformer causes the annoying non-linear distortion on the output signal due to the non-linear activation functions. The authors previously proposed a distortion-less neural network-based beamformer with the dual cost functions based on the directivity and spectral distortion. The beamformer was alternatively trained with either the directivity-based or spectral distortion-based cost functions. The proposed beamformer was slightly superior to the conventional non-linear beamformer in the viewpoints of the signal-to-noise ration and spectral distortion. In this paper, the number of the network layers is increased from three to four and five, and the training scheme is reconsidered in training the neural network. It is found that the neural network-based beamformer with five layers achieves the highest performance. Concerning the training scheme, it is necessary to carefully investigate the relationship between the training order with the dual cost functions and the number of iteration. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Alternative training / Beamformer / Neural network / Non-linear distortion |
| Paper # | EA2018-64,EMM2018-64 |
| Date of Issue | 2018-11-14 (EA, EMM) |
| Conference Information | |
| Committee | EA / ASJ-H / EMM / IPSJ-MUS |
|---|---|
| Conference Date | 2018/11/21(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Hotel Koshuen |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | [Beginners Session] Engineering/Electro Acoustics, Psychological and Physiological Acoustics, Music and Computer, Content Processing, Digital Watermarking, and Related Topics |
| Chair | Suehiro Shimauchi(Kanazawa Inst. of Tech.) / 小澤 賢司(山梨大) / Keiichi Iwamura(TUC) / 吉井 和佳(京大) |
| Vice Chair | Kenichi Furuya(Oita Univ.) / Kanji Watanabe(Akita Pref. Univ.) / 中川 誠司(千葉大) / Minoru Kuribayashi(Okayama Univ.) / Tetsuya Kojima(NIT,Tokyo College) |
| Secretary | Kenichi Furuya(Shizuoka Inst. of Science and Tech.) / Kanji Watanabe(NHK) / 中川 誠司(尚絅大) / Minoru Kuribayashi(北陸先端大) / Tetsuya Kojima(千葉大) / (NIT, Tokyo) |
| Assistant | Keisuke Imoto(Ritsumeikan Univ.) / Daisuke Morikawa(Toyama Pref Univ.) / / Hiroko Akiyama(NIT, Nagano College) / キタヒロ カネダ(CANON) |
| Paper Information | |
| Registration To | Technical Committee on Engineering Acoustics / Auditory Research Meeting / Technical Committee on Enriched MultiMedia / Special Interest Group on Music and Computer |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Improvement of speech distortion caused by multi-layered neural network-based beamformer with alternative training |
| Sub Title (in English) | |
| Keyword(1) | Alternative training |
| Keyword(2) | Beamformer |
| Keyword(3) | Neural network |
| Keyword(4) | Non-linear distortion |
| 1st Author's Name | Kouhei Eguchi |
| 1st Author's Affiliation | Kyushu Institute of Technology(KIT) |
| 2nd Author's Name | Mitsunori Mizumachi |
| 2nd Author's Affiliation | Kyushu Institute of Technology(KIT) |
| Date | 2018-11-21 |
| Paper # | EA2018-64,EMM2018-64 |
| Volume (vol) | vol.118 |
| Number (no) | EA-312,EMM-313 |
| Page | pp.pp.7-12(EA), pp.7-12(EMM), |
| #Pages | 6 |
| Date of Issue | 2018-11-14 (EA, EMM) |