Cross-linguistic universality in speech emotion recognition: Comparing multilingual and monolingual computational speech models

Human listeners can recognize emotions from speech alone, even in unfamiliar languages (universality), while performing better in familiar languages even without lexical cues (in-group advantage). However, whether computational speech models exhibit these abilities remains unclear—let alone how multilingual training influences them, a question still understudied in human research as well. We explored these questions by comparing a multilingual Whisper model to its English-only counterpart across seven acted-speech corpora: three English and four non-English languages seen by the multilingual model (Canadian French, Japanese, Thai, Greek). Emotion classification results showed strong universality and weak in-group advantage compared to past human studies—the monolingual model’s high accuracy on non-English speech (78.95%) suggests strong universality, while its lower accuracy than the multilingual model (81.37%) reflects weak in-group advantage. Notably, the two models performed similarly on English (monolingual: 76.49%, multilingual: 77.73%), suggesting that multilingual exposure does not greatly improve emotion recognition even in familiar languages. Layer-wise probing showed performance remaining high across the network, indicating that the systems seem to represent emotion across layers. These findings suggest that large speech models partially mirror human abilities in emotional speech recognition and encode a surprisingly universal representation of emotion without language-specific fine-tuning.