@techreport{oai:ipsj.ixsq.nii.ac.jp:00226434,
 author = {下西, 莞太 and 福森, 隆寛 and 山下, 洋一 and Kanta, Shimonishi and Takahiro, Fukumori and Yoichi, Yamashita},
 issue = {63},
 month = {Jun},
 note = {本報告では，環境音分離において分離音の歪みだけでなく再混合音と入力混合音の誤差を考慮することで，分離性能の向上を目指したものである．環境音分離において理想的な分離が可能な場合，分離音を再混合させると入力混合音に戻ることが期待されるが，実際は分離歪みにより歪みが発生した再混合音となる可能性がある．従来の音源分離モデルの学習では，分離音と正解音源との誤差のみを損失関数に用いる．従来の損失関数に加え，再混合音と入力混合音の差である Remix loss を導入することで，分離音の誤差だけでなく再混合音と入力混合音との再混合誤差も考慮しながら学習が可能となる．評価実験では，Remix loss を L1 ノルムや SI-SDR loss それぞれ用い，学習重みパラメータを変更するなど様々な条件下で性能を比較した．実験結果より，いくつかの条件で SI-SDRi が向上し，最大で 0.5dB の分離性能向上がみられた．また，適切な学習重みパラメータを設定しなければ効果的でないこともわかった．, This report aims to improve the performance of environmental sound separation by considering not only the separated sound’s distortion but also the error between the remixed sound and the input sound. When ideal separation is achieved in ambient sound separation, remixing the separated sounds is expected to return to the input sound mixture. Still, the remixed sound may be distorted due to separation distortion. In conventional sound separation model learning, only the error between the separated sound and the correct source is used as the loss function. By introducing the remix loss, which is the difference between the remixed and input mixtures, in addition to the conventional loss function, it is possible to learn while considering the error of the separated sound and the remix error between the remixed sound and input mixtures. In the evaluation experiments, we compared the performance under various conditions, such as using the L1 norm and SI-SDR loss for the remix loss and changing the learning weight parameters. The experimental results showed that SI-SDRi improved under several conditions, with a maximum improvement of 0.5 dB in separation performance. It was also found that the performance is only effective if the appropriate learning weight parameters are set.},
 title = {分離歪みと再混合誤差を考慮した環境音分離},
 year = {2023}
}