Speech Emotion Recognition via CNN-LSTM Self-Attention

The patent replaces conventional machine learning methods (HMM, SVM) with deep learning neural networks (CNN, LSTM). This substitution transitions from traditional mechanical/mathematical processing to neural network-based processing, achieving higher emotion recognition accuracy while accepting increased system complexity. The CNN-LSTM architecture specifically substitutes previous methods to capture both spatial and temporal features in speech data.

The patent employs a composite neural network architecture combining CNN and LSTM layers. The CNN component extracts local spectral features while the LSTM component captures temporal dependencies. This composite structure integrates the strengths of both network types to achieve superior emotion recognition performance compared to using either method alone.

The patent applies local quality by enabling different parts of the speech signal to be processed with different importance weights. The attention mechanism identifies and emphasizes local regions containing emotion-critical information while suppressing less relevant segments. This allows the system to focus computational resources on emotionally significant portions of the utterance.

The attention mechanism serves as an intermediary between the LSTM output and the final classification layer. It selectively weights different time steps in the sequence, acting as a mediator that determines which portions of the processed speech signal should contribute most to the emotion classification decision.

The patent replaces manual low-level descriptor (LLD) feature extraction with automatic deep learning-based feature learning. The CNN layers automatically learn and extract relevant spectral features from raw speech spectrograms, eliminating the need for manual feature engineering. This substitution achieves superior accuracy while increasing computational complexity.

The neural network performs self-service by automatically learning optimal feature representations from the input speech data without human intervention. The CNN architecture autonomously identifies and extracts discriminative features during training, replacing the manual feature extraction process that previously required expert knowledge and manual tuning.