Scholarlink

<p>Speech Emotion Recognition (SER) plays a crucial role in advancing human–computer interaction by enabling systems to interpret and respond to human emotions more naturally. However, existing SER approaches often struggle to capture complex emotional variations and maintain robustness across different speakers and acoustic conditions. To address these limitations, this paper introduces an enhanced SER framework that integrates the Signed Cumulative Distribution Transform with Progressive Graph Convolutional Networks (SER-PGCN). The primary objective of this study is to develop an effective and generalizable SER model capable of learning deeper emotional representations while preserving the intrinsic statistical and temporal characteristics of speech signals. The proposed framework begins with an Adaptive Two-Stage Unscented Kalman Filter (ATSUKF) for signal normalization and noise reduction. Emotional descriptors such as MFCCs, chroma features, spectral contrast, zero-crossing rate, pitch, and energy are then transformed using a novel SCDT-based feature extraction pipeline, which effectively captures distribution-aware emotional patterns. These transformed features are subsequently modelled as graphs and processed through a Progressive Graph Convolutional Network that learns increasingly enriched relational embeddings across layers, forming the core innovation of the proposed approach. Emotions including calm, happy, sad, angry, fearful, disgust, and surprised are classified using the final PGCN outputs. The SER-PGCN model is implemented in Python and evaluated using precision, recall, specificity, accuracy, F1-score, and computation time. Experimental results demonstrate that the proposed method achieves an accuracy of 95.86%, outperforming state-of-the-art models such as RSER-CENN, SER-MFGCN, and SER-ACO-SVM. These findings confirm that the combination of SCDT with progressive graph learning significantly enhances accuracy, robustness, and computational efficiency in speech emotion recognition systems.</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

Enhanced speech emotion detection via Signed Cumulative Distribution Transform and Progressive Graph Convolutional Networks

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Seema Kedar,
  • Pradnya Thakre,
  • Archana Jadhav,
  • Geetanjali S. Mate,
  • Dipali Himmatrao Patil

摘要

Speech Emotion Recognition (SER) plays a crucial role in advancing human–computer interaction by enabling systems to interpret and respond to human emotions more naturally. However, existing SER approaches often struggle to capture complex emotional variations and maintain robustness across different speakers and acoustic conditions. To address these limitations, this paper introduces an enhanced SER framework that integrates the Signed Cumulative Distribution Transform with Progressive Graph Convolutional Networks (SER-PGCN). The primary objective of this study is to develop an effective and generalizable SER model capable of learning deeper emotional representations while preserving the intrinsic statistical and temporal characteristics of speech signals. The proposed framework begins with an Adaptive Two-Stage Unscented Kalman Filter (ATSUKF) for signal normalization and noise reduction. Emotional descriptors such as MFCCs, chroma features, spectral contrast, zero-crossing rate, pitch, and energy are then transformed using a novel SCDT-based feature extraction pipeline, which effectively captures distribution-aware emotional patterns. These transformed features are subsequently modelled as graphs and processed through a Progressive Graph Convolutional Network that learns increasingly enriched relational embeddings across layers, forming the core innovation of the proposed approach. Emotions including calm, happy, sad, angry, fearful, disgust, and surprised are classified using the final PGCN outputs. The SER-PGCN model is implemented in Python and evaluated using precision, recall, specificity, accuracy, F1-score, and computation time. Experimental results demonstrate that the proposed method achieves an accuracy of 95.86%, outperforming state-of-the-art models such as RSER-CENN, SER-MFGCN, and SER-ACO-SVM. These findings confirm that the combination of SCDT with progressive graph learning significantly enhances accuracy, robustness, and computational efficiency in speech emotion recognition systems.