Cross-Lingual Speech Emotion Recognition: Humans vs. Self-Supervised Models

TL;DR

The paper investigates cross-lingual speech emotion recognition by juxtaposing self-supervised learning (SSL) models with human performance across monolingual, cross-lingual, and transfer-learning regimes. It combines a layer-wise analysis of SSL representations with parameter-efficient fine-tuning (PEFT) and extends evaluation to fine-grained speech emotion diarization (SED) and dialect effects, including a Chinese Tianjin dialect. Across multilingual datasets (Mandarin, German, English) and a non-public TJ dialect corpus, SSL models demonstrate strong monolingual SER but suffer in cross-lingual settings, with transfer learning mitigating the gap; humans generally outperform models in cross-lingual and segment-level tasks, and dialects introduce notable perceptual differences. The work provides new insights into how SSL models perceive emotion across languages, the effectiveness of PEFT for cross-lingual adaptation, and the persistent role of dialect in emotion perception, informing future cross-lingual SER development.

Abstract

Utilizing Self-Supervised Learning (SSL) models for Speech Emotion Recognition (SER) has proven effective, yet limited research has explored cross-lingual scenarios. This study presents a comparative analysis between human performance and SSL models, beginning with a layer-wise analysis and an exploration of parameter-efficient fine-tuning strategies in monolingual, cross-lingual, and transfer learning contexts. We further compare the SER ability of models and humans at both utterance- and segment-levels. Additionally, we investigate the impact of dialect on cross-lingual SER through human evaluation. Our findings reveal that models, with appropriate knowledge transfer, can adapt to the target language and achieve performance comparable to native speakers. We also demonstrate the significant effect of dialect on SER for individuals without prior linguistic and paralinguistic background. Moreover, both humans and models exhibit distinct behaviors across different emotions. These results offer new insights into the cross-lingual SER capabilities of SSL models, underscoring both their similarities to and differences from human emotion perception.

Figures (1)

• Figure 1: Layer-wise analysis of CN and DE models under monolingual, cross-lingual and transfer learning settings.