Cross-Corpus Speech Emotion Recognition Based on Dynamically Filtering Multistage Diffusion Model

In cross-corpus speech emotion recognition (SER) tasks, although generative methods have shown certain potential in alleviating the scarcity of target domain samples, the generated samples often lack precise alignment with the acousticsemantic features of the target domain. This results in pseudo samples of uneven quality and even introduces noise, thereby weakening the effectiveness of transfer learning. To address the problem of insufficient quality control over pseudo samples in existing methods, we propose a Dynamically Filtering Multi- Stage Diffusion Model (DFMDM). The proposed DFMDM model establishes a three-stage iterative optimization process of “Generation– Filtering–Enhancement.” In the first stage, the conditional diffusion model is guided by an adversarial domain classifier to generate pseudo samples with target domain characteristics. In the second stage, a dynamic filtering module is presented to eliminate low-quality pseudo samples that significantly deviate from the target domain features. In the third stage, the filtered pseudo samples are mixed with source domain samples and reused to regenerate target-style pseudo samples. In this way, a progressive diffusion-based enhancement loop is constructed, which improves sample quality through iterative generation and dynamic adjustment of similarity thresholds. Experimental results conducted on CASIA, EmoDB, eNTERFACE and IEMOCAP demonstrate that the proposed DFMDM model achieves significant improvements over existing state-of-the-art (SOTA) approaches, fully showcasing its application potential and practical value in cross-corpus speech emotion recognition.