TL;DR
This paper introduces EmoDB, a comprehensive database of acted German emotional speech, featuring 10 actors performing 10 sentences across 7 emotions. This resource addresses the need for controlled, high-quality emotional speech data for research in human-machine interfaces and speech analysis, enabling detailed phonetic and prosodic investigations.
What Problem Does This Paper Address?
Investigating emotional expression in speech is crucial for developing more natural and intuitive human-machine interfaces. However, acquiring realistic emotional speech data for research presents significant challenges. "Real-life" emotional speech is often fleeting, difficult to capture consistently, and ethically problematic to record in intense emotional states. Conversely, purely synthetic or nonsense speech can lead to unnatural or exaggerated emotional performances. This paper tackles this dilemma by creating a controlled database of acted emotional speech that aims to balance naturalness with experimental rigor. The goal is to provide a standardized, high-quality dataset that allows for reliable analysis of emotional speech characteristics.
Methods
The authors developed EmoDB by recording 10 actors (5 male, 5 female) performing 10 German utterances each, simulating 7 distinct emotions: neutral, anger, fear, joy, sadness, disgust, and boredom. The recordings were made in an anechoic chamber using high-quality equipment, capturing both audio and electro-glottograms. To ensure comparability and control, all actors spoke the same verbal content, and the actors were instructed to use the Stanislavski method (self-induction by recalling past emotional experiences) to evoke emotions. The database was then rigorously evaluated through perception tests to assess emotion recognizability and naturalness. Utterances meeting specific thresholds for recognition (>80%) and naturalness (>60%) were selected and phonetically transcribed with detailed markers for voice quality, phonatory, and articulatory settings.
Datasets
The paper introduces the EmoDB (Expressive Speech Database). This newly created benchmark consists of approximately 800 German emotional speech utterances. It differs from existing datasets by focusing on acted emotional speech with a strong emphasis on phonetic detail and controlled experimental conditions. The database includes audio recordings, electro-glottograms, and detailed phonetic transcriptions.
Key Insights and Ideas
- The "Acted vs. Real" Dilemma: The paper highlights the inherent trade-offs in emotional speech research. While "real" emotions are ideal for naturalness, they are difficult to obtain and control. Acted emotions offer control but risk artificiality. EmoDB attempts to bridge this gap by using trained actors and a specific elicitation method (Stanislavski) to achieve a balance.
- Controlled Environment for Detailed Analysis: Recording in an anechoic chamber with high-quality equipment and capturing electro-glottograms was crucial for enabling detailed phonetic and prosodic analyses, including inverse filtering and precise articulatory feature extraction.
- Perception-Driven Selection: The rigorous evaluation process, where listeners judged recognizability and naturalness, was key to filtering the data. This ensures that the final database contains utterances that are not only emotionally expressive but also perceived as reasonably natural by human listeners.
- Phonetic Richness: The selection of everyday sentences, chosen for their phonetic content and potential for articulatory reduction, makes EmoDB valuable for studying subtle phonetic variations associated with emotional states.
Key Results
- High Recognition Rates: The perception tests showed that utterances in EmoDB achieved high recognition rates for emotions, with "anger" being the most recognizable (96.9%) and "disgust" the least (79.6%), though still above the 80% threshold for inclusion.
- Naturalness Threshold Met: Approximately 500 out of 800 recorded utterances met the criteria of being recognized better than 80% and judged as natural by more than 60% of listeners, forming the core of the database.
- Comprehensive Annotation: The selected utterances were meticulously labeled with narrow phonetic transcriptions (using SAMPA), including detailed markers for voice quality and articulatory settings, as well as syllable segmentation and stress markings.
Limitations and Future Work
- Acted Nature: The authors acknowledge that the database consists of acted emotions, which may not perfectly replicate the nuances of spontaneous emotional speech. While efforts were made to enhance naturalness, it remains a limitation.
- Limited Text Material: While the 10 sentences were chosen for their phonetic properties and interpretability, a larger and more diverse set of texts could further enhance the database's utility.
- Potential for Overacting: Despite instructions, there's an inherent risk of actors overacting, especially when trying to convey strong emotions.
- Future Work: The paper suggests that the database has already served as a basis for numerous studies on prosodic and articulatory features, and its availability online facilitates further research into speech synthesis, emotion recognition, and the acoustic correlates of emotional expression.
Summary updated