Scholarlink

Speech emotion recognition, with its wide range of applications in robotics, customer service, and HCI (human-computer interaction), is being continuously improved using new models, features, and feature extraction mechanisms. This study proposes 3 deep learning models: 1D CNN, 1D CNN combined with LSTM, and 1D CNN combined with BiLSTM, these models exhibit remarkable performance on a merged dataset comprising RAVDESS, TESS, CREMA-D, and SAVEE emotional speech corpora. Furthermore, they demonstrate strong performance when evaluated on individual databases. For reference, experiments using conventional machine learning algorithms, such as SVM, logistic regression, and ensemble methods are also performed on the datasets. The proposed deep learning models leverage a fusion of acoustic features including MFCC, RMSE, and ZCR, for training. Augmentation techniques like noise addition and pitch scaling were employed to enhance the models’ robustness. Notably, the proposed 1D CNN model achieves an impressive accuracy of 96% on the combined datasets, outperforming prior studies on merged datasets.

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 Classification with Acoustic Features Using 1D CNN and LSTM Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yesha Vyas,
  • Shofiya Bootwala,
  • Neha Soni

摘要

Speech emotion recognition, with its wide range of applications in robotics, customer service, and HCI (human-computer interaction), is being continuously improved using new models, features, and feature extraction mechanisms. This study proposes 3 deep learning models: 1D CNN, 1D CNN combined with LSTM, and 1D CNN combined with BiLSTM, these models exhibit remarkable performance on a merged dataset comprising RAVDESS, TESS, CREMA-D, and SAVEE emotional speech corpora. Furthermore, they demonstrate strong performance when evaluated on individual databases. For reference, experiments using conventional machine learning algorithms, such as SVM, logistic regression, and ensemble methods are also performed on the datasets. The proposed deep learning models leverage a fusion of acoustic features including MFCC, RMSE, and ZCR, for training. Augmentation techniques like noise addition and pitch scaling were employed to enhance the models’ robustness. Notably, the proposed 1D CNN model achieves an impressive accuracy of 96% on the combined datasets, outperforming prior studies on merged datasets.