EmoTale: An Enacted Speech-emotion Dataset in Danish

2025-08-20

AI Summary

To address the scarcity of modern emotional speech datasets for low-resource languages such as Danish, this work introduces EmoTale—the first bilingual (Danish/English) acted emotional speech corpus—filling a longstanding gap in Danish emotional speech resources since the 1997 Danish Emotional Speech (DES) corpus. EmoTale encompasses six basic emotions, recorded by professional actors and annotated with fine-grained metadata, enabling cross-lingual affective modeling and comparative analysis. We evaluate emotion recognition performance using self-supervised speech representations (SSLME embeddings) and openSMILE acoustic features under leave-one-speaker-out cross-validation; the Danish subset achieves an unweighted average recall (UAR) of 64.1%, matching DES’s performance and confirming the corpus’s validity and utility. This contribution provides both a high-quality benchmark dataset and a methodological framework for emotion recognition in low-resource languages.

Technology Category

• Natural Language Processing: Lexical Semantics and Morphology
• Humans and AI: Emotional Intelligence
• Cognitive Modeling & Cognitive Systems: Affective Computing

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

While multiple emotional speech corpora exist for commonly spoken languages, there is a lack of functional datasets for smaller (spoken) languages, such as Danish. To our knowledge, Danish Emotional Speech (DES), published in 1997, is the only other database of Danish emotional speech. We present EmoTale; a corpus comprising Danish and English speech recordings with their associated enacted emotion annotations. We demonstrate the validity of the dataset by investigating and presenting its predictive power using speech emotion recognition (SER) models. We develop SER models for EmoTale and the reference datasets using self-supervised speech model (SSLM) embeddings and the openSMILE feature extractor. We find the embeddings superior to the hand-crafted features. The best model achieves an unweighted average recall (UAR) of 64.1% on the EmoTale corpus using leave-one-speaker-out cross-validation, comparable to the performance on DES.

Problem

• Lack of emotional speech datasets for Danish language
• Need for validated emotional speech recognition models
• Comparison of self-supervised versus hand-crafted feature performance

Innovation

• Developed Danish-English emotional speech dataset
• Used self-supervised speech model embeddings
• Achieved 64.1% UAR with cross-validation