Speech Enhancement Based on Cyclegan with Noise-informed Training

TL;DR

This paper introduces a noise-informed training method for CycleGAN-based speech enhancement, improving its ability to generalize and perform better in noisy environments by incorporating target domain information during training.

Contribution

The paper proposes a novel noise-informed training approach that enhances CycleGAN speech enhancement systems by better utilizing target domain information during clean-to-noisy conversion.

Findings

NIT improves CycleGAN speech enhancement performance

Enhanced generalization capability demonstrated

Significant margin of improvement in experiments

Abstract

Cycle-consistent generative adversarial networks (CycleGAN) were successfully applied to speech enhancement (SE) tasks with unpaired noisy-clean training data. The CycleGAN SE system adopted two generators and two discriminators trained with losses from noisy-to-clean and clean-to-noisy conversions. CycleGAN showed promising results for numerous SE tasks. Herein, we investigate a potential limitation of the clean-to-noisy conversion part and propose a novel noise-informed training (NIT) approach to improve the performance of the original CycleGAN SE system. The main idea of the NIT approach is to incorporate target domain information for clean-to-noisy conversion to facilitate a better training procedure. The experimental results confirmed that the proposed NIT approach improved the generalization capability of the original CycleGAN SE system with a notable margin.