End-to-end speech emotion recognition using a novel context-stacking dilated convolution neural network.

Area of Science:

• Artificial Intelligence
• Signal Processing
• Computational Linguistics

Background:

• Speech emotion recognition (SER) requires models that can learn long temporal dependencies due to the slow dynamics of emotional expression in speech.
• Recurrent Neural Networks (RNNs) are commonly used but suffer from parallelization limitations.
• Existing SER systems often rely on hand-crafted features, which may not capture all relevant information.

Purpose of the Study:

• To propose a novel end-to-end neural network architecture for speech emotion recognition.
• To develop a model capable of learning long-term temporal dependencies efficiently.
• To compare the proposed model's performance and parameter efficiency against state-of-the-art SER systems.

Main Methods:

• A novel end-to-end neural network architecture utilizing dilated causal convolution with context stacking.
• The model employs parallelizable layers, avoiding RNN limitations.
• Receptive field size is maximized to the input sequence length with controlled computational cost.

Main Results:

• The proposed model significantly improves SER performance in both regression and classification tasks.
• It achieves superior results using only 1/3 of the parameters compared to a state-of-the-art model.
• The model effectively learns intermediate embeddings that preserve speech emotion information.

Conclusions:

• The proposed dilated causal convolution network with context stacking offers an efficient and effective alternative to RNNs for SER.
• End-to-end learning with this architecture surpasses traditional methods using hand-crafted features.
• The model demonstrates strong potential for real-world applications requiring accurate speech emotion analysis.