Scholarlink

<p>Speech Emotion Recognition (SER) is an important challenge in the realm of Human Computer Interaction (HCI). Its successful implementation can be very useful in many applications like online teaching, call centers, therapy sessions and more. Deep Learning, a field within machine learning, has been effectively used by researchers for training SER models with high accuracy. In our work, experimental comparisons of three hybrid architectures have been explored. In each hybrid model, three types of Convolutional Neural Networks (CNNs), namely 1D, 2D, and 3D CNNs, have been combined, based on averaging (Model A), parallel merging (Model B) and sequential integration (Model C). The three hybrid models have been designed to leverage the fusion of acoustic features – ZCR, RMS, Chroma-STF, MFCCs and Mel spectrograms extracted from speech signals. A systematic comparison of the three hybrid CNNs for selected datasets was conducted. Among the datasets there were two semi-natural datasets - Interactive Emotional Dyadic Motion Capture Database (IEMOCAP) and Database of Elicited Mood in Speech (DEMoS) and three acted datasets - Toronto emotional speech set (TESS), Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and Berlin Emotional Database (EMODB). Model A demonstrated overall superiority in terms of accuracy, making it suitable for developing real-world applications.</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

Comparison of three hybrid architectures using 1D, 2D, and 3D CNNs for speech emotion recognition

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Aparna Vyakaranam,
  • Tomas Maul,
  • Bavani Ramayah

摘要

Speech Emotion Recognition (SER) is an important challenge in the realm of Human Computer Interaction (HCI). Its successful implementation can be very useful in many applications like online teaching, call centers, therapy sessions and more. Deep Learning, a field within machine learning, has been effectively used by researchers for training SER models with high accuracy. In our work, experimental comparisons of three hybrid architectures have been explored. In each hybrid model, three types of Convolutional Neural Networks (CNNs), namely 1D, 2D, and 3D CNNs, have been combined, based on averaging (Model A), parallel merging (Model B) and sequential integration (Model C). The three hybrid models have been designed to leverage the fusion of acoustic features – ZCR, RMS, Chroma-STF, MFCCs and Mel spectrograms extracted from speech signals. A systematic comparison of the three hybrid CNNs for selected datasets was conducted. Among the datasets there were two semi-natural datasets - Interactive Emotional Dyadic Motion Capture Database (IEMOCAP) and Database of Elicited Mood in Speech (DEMoS) and three acted datasets - Toronto emotional speech set (TESS), Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and Berlin Emotional Database (EMODB). Model A demonstrated overall superiority in terms of accuracy, making it suitable for developing real-world applications.