Source research

Improving Cross-Corpus Speech Emotion Recognition with Adversarial Discriminative Domain Generalization (ADDoG)

John Gideon, Melvin G McInnis, Emily Mower Provost · 2019 · 93 citationsRead the paper

Automatic speech emotion recognition provides computers with critical context to enable user understanding. While methods trained and tested within the same dataset have been shown successful, they often fail when applied to unseen datasets. To address this, recent work has focused on adversarial methods to find more generalized representations of emotional speech. However, many of these methods have issues converging, and only involve datasets collected in laboratory conditions. In this paper, we introduce Adversarial Discriminative Domain Generalization (ADDoG), which follows an easier to train "meet in the middle" approach. The model iteratively moves representations learned for each dataset closer to one another, improving cross-dataset generalization. We also introduce Multiclass ADDoG, or MADDoG, which is able to extend the proposed method to more than two datasets, simultaneously. Our results show consistent convergence for the introduced methods, with significantly improved results when not using labels from the target dataset. We also show how, in most cases, ADDoG and MADDoG can be used to improve upon baseline state-of-the-art methods when target dataset labels are added and in-the-wild data are considered. Even though our experiments focus on cross-corpus speech emotion, these methods could be used to remove unwanted factors of variation in other settings.

What the full text shows

Working prototype

ADDoG significantly outperforms CNN baseline when no target labels are available for laboratory datasets. — 0.474±0.009 UAR (MSP-Improv to IEMOCAP) vs 0.439±0.022 UAR; 0.444±0.007 UAR (IEMOCAP to MSP-Improv) vs 0.432±0.012 UAR (Laboratory datasets (IEMOCAP, MSP-Improv), 0 target labels)
ADDoG provides significant improvement over baseline when training on lab data and testing on in-the-wild data with few labels. — Significant improvement for ≤ 800 labelled target utterances (Lab (IEMOCAP/MSP-Improv) to in-the-wild (PRIORI Emotion))
MADDoG outperforms other methods when integrating multiple source datasets for in-the-wild testing. — Significantly better than others in most cases except ADDoG with 200 samples (IEMOCAP and MSP-Improv to PRIORI Emotion)
Specialization (SP) outperforms ADDoG when a large amount of target labels are available and datasets are very different. — '' (High volume of target labels, high domain mismatch (e.g., Lab to In-the-wild))

Setup: Experiments used IEMOCAP, MSP-Improv (laboratory), and PRIORI Emotion (in-the-wild) datasets; models were trained for 30-50 epochs on GPUs (GTX 1080, Titan X).
How: The authors developed Adversarial Discriminative Domain Generalization (ADDoG) and its multiclass extension (MADDoG), which use a 'meet in the middle' adversarial approach to iteratively move dataset representations closer to each other. These were compared against a baseline CNN and a specialization (SP) model across various cross-corpus scenarios.
Limitations: Performance of generalized models (ADDoG) decreases relative to specialized models (SP) when target labels are abundant. · Performance drops significantly when moving from laboratory to in-the-wild data due to recording and elicitation mismatches.

Summarized by Gemma 4 from the open-access full text · 3 Oct 2026. Read it