Research on Data Augmentation and Deep Learning Algorithms for Speech Emotion Computing

Abstract

Speech is the most natural human interaction medium, which contains both linguistic semantic information and numerous non-semantic linguistic characteristics such as emotion, attitude and mood, making it a core research object in the fields of computational linguistics and human-computer interaction. Speech emotion computing aims to explore emotional features embedded in speech signals through computer algorithms to realize automatic recognition and classification of human emotions. It is an interdisciplinary research area integrating linguistics, signal processing and artificial intelligence. At present, mainstream research on speech emotion recognition generally faces problems including insufficient annotated corpora, single sample dimensions, poor model generalization ability and low recognition accuracy under low-resource conditions. This paper systematically conducts research from two major perspectives: data augmentation and deep learning algorithms. Firstly, it sorts out the computational linguistic theoretical basis of speech emotion computing and analyzes the linguistic implications of prosodic, spectral and temporal emotional features in speech signals. Secondly, it investigates traditional data augmentation and intelligent deep learning-based augmentation methods separately, and optimizes augmentation strategies according to the linguistic characteristics of oral speech corpora. In addition, a deep learning recognition model suitable for speech emotional features is constructed to improve the feature extraction capability and classification performance of the model. Finally, extensive comparative experiments are carried out to verify the joint optimization effect of the proposed data augmentation algorithms and deep learning models. Experimental results demonstrate that the improved multi-dimensional data augmentation method can effectively enrich speech corpora and address the shortage of low-resource annotated data. Combined with the optimized deep learning model, the proposed method significantly improves the accuracy and stability of speech emotion recognition, providing technical support for natural language interaction scenarios such as intelligent customer service, human-computer dialogue and emotion monitoring.