Seminar "Artificial intelligence to identify depression from audio information"

On May 29, a scientific seminar on the topic “Artificial intelligence to identify depression from audio information” was held at the Laboratory of Artificial Intelligence for Cognitive Sciences

The speaker was Anna Kazachkova (master's student at the HSE) under the supervision of Shalileh Soroosh (Ph.D. in Computer Science, Laboratory Head).

Annotation: 
Depression is a widespread psychiatric disorder, which can significantly deteriorate the quality of life. Automatic depression detection could be an accessible and reliable diagnostic tool, addressing the current issues in the mental disorders area. The purpose of this paper is to study how accurately depression can be predicted on a given dataset and what are the most sustainable models and data representations. The study focuses on problem formulations such as binary classification and abnormality detection. The exploited models included convolutional neural networks and the transformer, and they were either trained only on our dataset or employed in the form of pre-trained for the image classification instances. Additionally, a benchmark of classical machine learning algorithms for the Geneva Minimalistic Acoustic Parameter Set features was computed. In total, we derived the best average ROC-AUC value of 0.72 on the test, compared to the benchmark of 0.55. This best result was provided by fine-tuning InceptionV3 architecture under the one-plus-epsilon optimization algorithm.

Materials:  Presentation from the seminar