Scholarlink

Emotions are an essential part of human communication since they impact how we coexist with each other, and how we understand each other. The accurate identification and interpretation of those emotions are fundamental for the betterment of human-computer interaction and toward the development of more effective and sensitive systems. The application of SER technology is the identification and analysis of the emotional state communicated through the vocal expression of a person. It will include critical investigation of various speech attributes, such as tone, pitch, and rhythm, to determine the underlying emotions. The approach proposed gathers a wide variety of emotional speech samples from four separate datasets, namely RAVDESS, CREMA-D, SAVEE, and TESS, with the final objective of synthesizing an incredibly diverse and wide range of emotional speech samples to train on as well as evaluate with. Additional data augmentation techniques used to improve the robustness and generalization of the model. Using a hierarchical CNN-Transformer framework for the purpose of emotion identification. The model’s Transformer block is intended to simulate long-range dependencies in the sequence data, whilst the first convolutional layers concentrate on collecting local temporal aspects. The model can learn and include both local and global features as both essential for precise emotion recognition. The incorporation of four different datasets helps our model attain the accuracy of over 92%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Speech Emotion Recognition Using a CNN-Transformer Approach on Diverse Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nidhi Bansal,
  • V. Malathy,
  • R. Lakshmikanth,
  • Krishan Arora,
  • Himanshu Sharma,
  • Abdul Rehman,
  • Nishant Upadhyay

摘要

Emotions are an essential part of human communication since they impact how we coexist with each other, and how we understand each other. The accurate identification and interpretation of those emotions are fundamental for the betterment of human-computer interaction and toward the development of more effective and sensitive systems. The application of SER technology is the identification and analysis of the emotional state communicated through the vocal expression of a person. It will include critical investigation of various speech attributes, such as tone, pitch, and rhythm, to determine the underlying emotions. The approach proposed gathers a wide variety of emotional speech samples from four separate datasets, namely RAVDESS, CREMA-D, SAVEE, and TESS, with the final objective of synthesizing an incredibly diverse and wide range of emotional speech samples to train on as well as evaluate with. Additional data augmentation techniques used to improve the robustness and generalization of the model. Using a hierarchical CNN-Transformer framework for the purpose of emotion identification. The model’s Transformer block is intended to simulate long-range dependencies in the sequence data, whilst the first convolutional layers concentrate on collecting local temporal aspects. The model can learn and include both local and global features as both essential for precise emotion recognition. The incorporation of four different datasets helps our model attain the accuracy of over 92%.