[2Xin5] インタラクティブ1

[2Xin5-22] Recovering Spectrograms using Contextual Attention

〇Shunsuke HABARA¹, Yoshiaki KUROSAWA¹, Kazuya MERA¹, Toshiyuki TAKEZAWA¹ (1.Graduate School of Information Sciences, Hiroshima City University)

Keywords:Voice, Deep Learning, Inpainting

There is a growing trend towards implementing technologies that use deep neural networks to improve sound quality by signal denoising, and a system that converts voice quality in real-time for the online conference. In the field of computer vision, inpainting techniques based on deep neural networks have also been developed in recent years. In this paper, we focus on an inpainting technique with contextual attention to recover spectrograms. We apply a mask to the time direction of the spectrogram and examine whether the spectrogram can be recovered from the non-masked area. We propose a method to improve the accuracy of speech restoration by providing a gradient in the frequency direction to the spectrogram. As a result, our proposed method improved one of sound metrics: Mel-Cepstral Distortion. We also confirmed that the attention map improved attention in the frequency.