Diffusion-based Speech Enhancement Using Decorrelated Attention for Score Network

Kim KR, Lee GW, Kim HK. Diffusion-based Speech Enhancement Using Decorrelated Attention for Score Network. In: AES Convention 155. Audio Engineering Society; 2023. Paper 175. Available from: https://aes.org/publications/elibrary-page/?id=22329

@inproceedings{Kim2023_22329,
  author = {Kim, Kang Ryeol and Lee, Geon Woo and Kim, Hong Kook},
  title = {{Diffusion-based Speech Enhancement Using Decorrelated Attention for Score Network}},
  booktitle = {AES Convention 155},
  note = {Paper 175},
  year = {2023},
  month = oct,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=22329}
}

TY  - CPAPER
TI  - Diffusion-based Speech Enhancement Using Decorrelated Attention for Score Network
AU  - Kim, Kang Ryeol
AU  - Lee, Geon Woo
AU  - Kim, Hong Kook
T2  - AES Convention 155
M1  - Paper 175
PY  - 2023
DA  - 2023/10/06
UR  - https://aes.org/publications/elibrary-page/?id=22329
PB  - Audio Engineering Society
LA  - en
AB  - In this paper, we propose decorrelated attention (DeCA)-based score network approach for speech enhancement that leverages background noise for perturbing speech data and decorrelates it from noise data. Based on the fundamental principle of a diffusion-based generative model that removes uncorrelated noise from perturbed speech, the DeCA and the corresponding loss are utilized to facilitate the decorrelation of the noisy data from the perturbed speech. The DeCA mechanism calculates an attention matrix, which is then multiplied with noisy data to achieve decorrelation. To evaluate the performance, the WSJ0-CHiME3 dataset was used for training and testing. The proposed DeCA-based score network model showed an enhanced performance across all metrics. Specifically, the scale-invariant signal-to-interference ratio (SI-SIR) showed an increase of 0.5 dB when compared to the original score network model.
ER  -