Acted Emotional Speech Dynamic Database

Speech Emotion Recognition

Speech Emotion Recognition (SER) is the process of extracting emotional paralinguistic information from speech. It is a field with growing interest and potential applications in Human-Computer Interaction, content management, social interaction, and as an add-on module in Speech Recognition and Speech-To-Text systems.
Text-independent automated SER relies on the specific attributes of speech audio signals. In such an elusive task as SER, typically a data-driven approach is followed. This means that models are trained on data. As a consequence, the performance of SER models is inextricably linked to the quality and the organization of the provided dataset.

Acted Emotional Speech Dynamic Database – AESDD

Databases of emotional speech are divided into two main categories, the ones that contain utterances of acted emotional speech and the ones that contain spontaneous emotional speech. Both categories have benefits and limitations.

The Acted Emotional Speech Dynamic Database (AESDD) contains utterances of acted emotional speech in the Greek language.

The motive for the creation of the database was the absence of a publically available high-quality database for SER in Greek, a realization made during the research on an emotion-triggered lighting framework for theatrical performance [1]. The database utterances with five emotions: anger, disgust, fear, happiness, and sadness.

The first version of the database was created in collaboration with a group of professional actors, who showed vivid interest in the proposed framework. Dynamic (in AESDD) refers to the intention of constantly expanding the database through the contribution of actors and performers that are involved, or interested in the project. While the call for contribution addresses to actors, the SER models that are trained on the AESDD are not exclusively performance-oriented.

The first version of the AESDD was presented in [2].

In [4], subjective evaluation experiments were carried out on the database, to assess human accuracy in recognizing the intended emotion in AESDD utterances. The accuracy of human listeners was estimated at around 74%.

Publications

  1. Vryzas, N., Liatsou, A., Kotsakis, R., Dimoulas, C., & Kalliris, G. (2017, August). Augmenting Drama: A Speech Emotion-Controlled Stage Lighting Framework. In Proceedings of the 12th International Audio Mostly Conference on Augmented and Participatory Sound and Music Experiences (p. 8). ACM.
  2. Vryzas, N., Kotsakis, R., Liatsou, A., Dimoulas, C. A., & Kalliris, G. (2018). Speech emotion recognition for performance interaction. Journal of the Audio Engineering Society, 66(6), 457-467.
  3. Vryzas, N., Vrysis, L., Kotsakis, R., & Dimoulas, C. (2018, September). Speech emotion recognition adapted to multimodal semantic repositories. In 2018 13th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP) (pp. 31-35). IEEE.
  4. Vryzas, N., Matsiola, M., Kotsakis, R., Dimoulas, C., & Kalliris, G. (2018, September). Subjective Evaluation of a Speech Emotion Recognition Interaction Framework. In Proceedings of the Audio Mostly 2018 on Sound in Immersion and Emotion (p. 34). ACM.

Download

The last version of the AESDD, as well as tools and documentation on the way the database is organized, can be found in the following link:

Acted Emotional Speech Dynamic Database

If you use the AESDD for scientific research please cite [2] and [4].

Contact

If there are any questions or if you would like to contribute to the project, please contact Nikolaos Vryzas (nvryzas@auth.gr) or use the form below: