RCT: Random Consistency Training for Semi-supervised Sound Event Detection

TL;DR

This paper introduces Random Consistency Training (RCT), a semi-supervised learning strategy for sound event detection that enhances model stability and performance without additional annotation costs.

Contribution

The paper proposes a novel RCT strategy with self-consistency loss, hard mixup augmentation, and random augmentation schemes for improved semi-supervised sound event detection.

Findings

RCT outperforms existing semi-supervised strategies

Self-consistency loss stabilizes training

Hard mixup improves sound additive property handling

Abstract

Sound event detection (SED), as a core module of acoustic environmental analysis, suffers from the problem of data deficiency. The integration of semi-supervised learning (SSL) largely mitigates such problem while bringing no extra annotation budget. This paper researches on several core modules of SSL, and introduces a random consistency training (RCT) strategy. First, a self-consistency loss is proposed to fuse with the teacher-student model to stabilize the training. Second, a hard mixup data augmentation is proposed to account for the additive property of sounds. Third, a random augmentation scheme is applied to flexibly combine different types of data augmentations. Experiments show that the proposed strategy outperform other widely-used strategies.