Speech Enhancement Based on Deep Mixture of Distinguishing Experts

In this work, we propose a new strategy for deep mixture of experts (DMoE) based speech enhancement. DMoE system is difficult to train due to the specific network structure and the necessity of carefully designed pre-training methods to guarantee good performance. We propose using distinguishing dee...

Full description

Saved in:
Bibliographic Details
Main Authors: Jia, Xupeng, Li, Dongmei
Format: Conference Proceeding
Language:English
Subjects:
Convergence
Mathematical model
Signal to noise ratio
Spectrogram
Speech enhancement
Time-frequency analysis
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this work, we propose a new strategy for deep mixture of experts (DMoE) based speech enhancement. DMoE system is difficult to train due to the specific network structure and the necessity of carefully designed pre-training methods to guarantee good performance. We propose using distinguishing deep neural networks (DNNs) as experts, dealing with magnitude spectrogram and log-magnitude spectrogram respectively. The proposed method is compared with the state-of-art DMoE system utilizing hard expectation maximization (HEM) pre-training method. Speech enhancement experiments in 30 (5*6) noise and SNR conditions show the superiority of the proposed method over the baseline method. The average improvements obtained for matched conditions are 0.076 in perceptual evaluation of speech quality (PESQ), 1.824dB in segmental signal to noise ratio (segSNR) and 0.043 in short time objective intelligibility (STOI).
ISSN:2640-0103
DOI:10.1109/APSIPAASC47483.2019.9023048