@techreport{oai:ipsj.ixsq.nii.ac.jp:00209770,
 author = {Jeongwoo, Woo and Masato, Mimura and Kazuyoshi, Yoshii and Tatsuya, Kawahara and Jeongwoo, Woo and Masato, Mimura and Kazuyoshi, Yoshii and Tatsuya, Kawahara},
 issue = {32},
 month = {Feb},
 note = {Speech enhancement in the presence of background music is not so different from noise reduction if music is treated as just noise. However, music has definite characteristics which are made by human beings, unlike noise which can be any. In order to consider characteristics of background music instead of noise reduction, we introduce a generative adversarial network (GAN). We combine two multi-scale discriminators for speech and music with Conv-TasNet modified for speech enhancement. We train it jointly with SI-SDR and the GAN objective. Experimental evaluations through speech recognition demonstrate that the proposed model is improved from the baseline model. It is notable that the more music interference is large, the more the proposed method is effective. Comparing the spectrogram of enhanced speech by the proposed and baseline model demonstrate that the baseline model tends to cut off noise excessively, in contrast the proposed model reconstructs more faithfully., Speech enhancement in the presence of background music is not so different from noise reduction if music is treated as just noise. However, music has definite characteristics which are made by human beings, unlike noise which can be any. In order to consider characteristics of background music instead of noise reduction, we introduce a generative adversarial network (GAN). We combine two multi-scale discriminators for speech and music with Conv-TasNet modified for speech enhancement. We train it jointly with SI-SDR and the GAN objective. Experimental evaluations through speech recognition demonstrate that the proposed model is improved from the baseline model. It is notable that the more music interference is large, the more the proposed method is effective. Comparing the spectrogram of enhanced speech by the proposed and baseline model demonstrate that the baseline model tends to cut off noise excessively, in contrast the proposed model reconstructs more faithfully.},
 title = {Speech Enhancement in the Presence of Background Music Considering Speech and Music Characteristics},
 year = {2021}
}