Scholarlink

<p>In our increasingly connected world, understanding human emotions is vital for fostering meaningful social connections and enhancing communication. This article introduces an innovative approach to emotion recognition that leverages two distinct deep learning algorithms.The first algorithm employs a hybrid neural network combining MultiLayer Perceptron (1-D CNN-MLP), while the second integrates CNN technology with support vector machines (1-D CNN-SVM).To assess the effectiveness of our proposed system, we utilize data augmentation techniques with the Ryerson Multimedia Lab (RML) dataset, extracting features such as Mel Frequency Cepstral Coefficients, Zero Crossing Rate, Mel Spectrograms, Root Mean Square,and Chromagram from the enriched data. Our 1-D CNN-MLP model demonstrates impressive accuracy rates of 85.93%,84.63%, and 81.85% using the ADAM, RMSprop, and SGD optimization functions,respectively, surpassing previous benchmarks.Furthermore, by combining the strengths of CNN and SVM with three different kernels (linear, RBF,and polynomial),we aim to enhance the accuracy and robustness of emotion recognition. Our 1-D CNN-SVM model achieves the highest recognition rate of 82.59% using the RMSprop optimization function with the RBF kernel. This comparative analysis highlights the innovative aspects of our approach, showcasing the superior performance of our models in capturing the nuances of various emotional states.These advancements significantly enhance the effectiveness and performance of our system, pushing the boundaries of emotion recognition from speech.</p>

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

The influence of optimizer function for emotion detection from speech with data augmentation

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Hadhami Aouani,
  • Yassine Ben Ayed

摘要

In our increasingly connected world, understanding human emotions is vital for fostering meaningful social connections and enhancing communication. This article introduces an innovative approach to emotion recognition that leverages two distinct deep learning algorithms.The first algorithm employs a hybrid neural network combining MultiLayer Perceptron (1-D CNN-MLP), while the second integrates CNN technology with support vector machines (1-D CNN-SVM).To assess the effectiveness of our proposed system, we utilize data augmentation techniques with the Ryerson Multimedia Lab (RML) dataset, extracting features such as Mel Frequency Cepstral Coefficients, Zero Crossing Rate, Mel Spectrograms, Root Mean Square,and Chromagram from the enriched data. Our 1-D CNN-MLP model demonstrates impressive accuracy rates of 85.93%,84.63%, and 81.85% using the ADAM, RMSprop, and SGD optimization functions,respectively, surpassing previous benchmarks.Furthermore, by combining the strengths of CNN and SVM with three different kernels (linear, RBF,and polynomial),we aim to enhance the accuracy and robustness of emotion recognition. Our 1-D CNN-SVM model achieves the highest recognition rate of 82.59% using the RMSprop optimization function with the RBF kernel. This comparative analysis highlights the innovative aspects of our approach, showcasing the superior performance of our models in capturing the nuances of various emotional states.These advancements significantly enhance the effectiveness and performance of our system, pushing the boundaries of emotion recognition from speech.