Bidirectional parallel echo state network for speech emotion recognition

Area of Science:

• Artificial Intelligence
• Speech Processing
• Machine Learning

Background:

• Speech signals are crucial for human communication and human-computer interaction.
• Emotion recognition from speech is a significant research area for developing more intuitive machines.
• Existing methods often struggle with the complexity and nuances of speech emotion data.

Purpose of the Study:

• To propose a novel speech emotion recognition (SER) system.
• To enhance machine understanding of human emotions through speech.
• To improve the accuracy and robustness of SER systems.

Main Methods:

• Utilized multivariate time series handcrafted features from speech signals.
• Implemented a bidirectional echo state network (ESN) with two parallel reservoir layers.
• Employed sparse random projection for dimensionality reduction and sampling techniques to handle imbalanced datasets.

Main Results:

• The proposed parallel ESN model demonstrated superior performance in speaker-independent experiments.
• Achieved better results compared to single reservoir ESN models.
• Outperformed existing state-of-the-art speech emotion recognition studies on benchmark datasets (EMO-DB, SAVEE, RAVDESS, FAU Aibo).

Conclusions:

• The novel parallel ESN architecture effectively captures complex speech features for emotion recognition.
• The proposed SER system offers a significant advancement in the field.
• This approach holds promise for more sophisticated human-computer interaction systems.