Scholarlink

<p>Emotion recognition using speech signals plays a crucial role in enhancing user interactions across various multimedia applications, including entertainment, healthcare, marketing, and education. Achieving high recognition accuracy while minimizing model complexity for real-time deployment on devices with limited computational resources presents a significant challenge. This challenge is addressed by the introduction of a novel, resource-efficient emotion recognition model: the Prosodic and Spectral, Alternating 1D Convolutional Model with Fixed Kernels and Incremental Filters (PS-AC1D-FIF). The key innovation lies in the combination of prosodic and spectral features, alternating 1D convolutions with batch normalization, fixed kernel sizes, and incremental filters without pooling. These components collectively optimize computational efficiency while maintaining recognition accuracy. The proposed method is evaluated using the EMO-DB and RAVDESS speech datasets, achieving accuracy scores of 87.85% and 81.12%, respectively. A 5-fold cross-validation procedure is employed, with performance assessed using standard metrics, including accuracy, precision, recall, F1 score, and Unweighted Accuracy (UWA). A comparison with state-of-the-art methods reveals a significant reduction in time and parameter complexity, with the EMO-DB dataset processed in 240 seconds with 529,207 parameters, and the RAVDESS dataset in 572 seconds with 563,908 parameters—substantially lower than prior approaches.</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

A lightweight architecture for speech emotion recognition using spectral & prosodic features with alternating convolutional layers and batch normalizations, fixed kernels, and incremental filters

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Sooma Heydari,
  • Seyed Mansour Shahidi

摘要

Emotion recognition using speech signals plays a crucial role in enhancing user interactions across various multimedia applications, including entertainment, healthcare, marketing, and education. Achieving high recognition accuracy while minimizing model complexity for real-time deployment on devices with limited computational resources presents a significant challenge. This challenge is addressed by the introduction of a novel, resource-efficient emotion recognition model: the Prosodic and Spectral, Alternating 1D Convolutional Model with Fixed Kernels and Incremental Filters (PS-AC1D-FIF). The key innovation lies in the combination of prosodic and spectral features, alternating 1D convolutions with batch normalization, fixed kernel sizes, and incremental filters without pooling. These components collectively optimize computational efficiency while maintaining recognition accuracy. The proposed method is evaluated using the EMO-DB and RAVDESS speech datasets, achieving accuracy scores of 87.85% and 81.12%, respectively. A 5-fold cross-validation procedure is employed, with performance assessed using standard metrics, including accuracy, precision, recall, F1 score, and Unweighted Accuracy (UWA). A comparison with state-of-the-art methods reveals a significant reduction in time and parameter complexity, with the EMO-DB dataset processed in 240 seconds with 529,207 parameters, and the RAVDESS dataset in 572 seconds with 563,908 parameters—substantially lower than prior approaches.