Scholarlink

Emotion intelligence is playing a crucial role in identifying the mental and physical health of humans. Corporates, government organizations, and schools are concentrating on mental health for better working environment. Automization of identifying emotions will enable the process of measuring the emotional intelligence. Speech Emotion Recognition (SER) is the best way to identify human emotions. In this paper, MFCC features from the non-linguistic details of speech are extracted for training the Stacking Ensemble Learner. Advanced neural networks, viz., 1D CNN and CapsuleNets are stacked as base learners, and multilayer perceptron is used at metalayer for decision-making. The experiments are conducted on two benchmark datasets RAVDESS and IEMOCAP. The classification accuracies achieved are 91.57% and 91.81%, respectively. Precision, recall, and F1-scores are also evaluated, and the results obtained reflect the proficiency of the proposed model over the existing model. Also, the classification of happy emotion, which is very crucial in SER, has encouraging results.

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

Stacking Ensemble Learning with 1D CNN and CapsuleNet for Speech Emotion Recognition

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yalamanchili Bhanusree,
  • B. Janardhana Rao,
  • K. Revathi,
  • Venkata Krishna Odugu,
  • Satish Bojjawar

摘要

Emotion intelligence is playing a crucial role in identifying the mental and physical health of humans. Corporates, government organizations, and schools are concentrating on mental health for better working environment. Automization of identifying emotions will enable the process of measuring the emotional intelligence. Speech Emotion Recognition (SER) is the best way to identify human emotions. In this paper, MFCC features from the non-linguistic details of speech are extracted for training the Stacking Ensemble Learner. Advanced neural networks, viz., 1D CNN and CapsuleNets are stacked as base learners, and multilayer perceptron is used at metalayer for decision-making. The experiments are conducted on two benchmark datasets RAVDESS and IEMOCAP. The classification accuracies achieved are 91.57% and 91.81%, respectively. Precision, recall, and F1-scores are also evaluated, and the results obtained reflect the proficiency of the proposed model over the existing model. Also, the classification of happy emotion, which is very crucial in SER, has encouraging results.