Scholarlink

<p>Speech Emotion Recognition (SER) has emerged as a crucial component in enhancing machine-user communication, particularly in conversational AI platforms and mental health monitoring systems. Despite the recent progress enabled by deep learning and pretrained speech models, existing SER systems still struggle with generalizability, temporal context preservation, and limited interpretability of acoustic features. This paper introduces Wav2TP, a novel SER framework that leverages temporal pooling techniques on contextualized embeddings extracted from the transformer-based Wav2Vec2.0 network. The proposed model incorporates statistical pooling strategies which are mean, standard deviation, 25th and 75th percentiles combined with Principal Component Analysis (PCA) to reduce redundancy and enhance the feature representation. A Multi-Layer Perceptron (MLP) classifier was subsequently used for emotion classification. The proposed method was then evaluated on the EmoDB dataset under four different configurations. The baseline model using mean pooling achieved an accuracy of 88.79%, while adding attention layers and additional pooling statistics progressively improved performance. The best-performing configuration achieved a 94.39% test accuracy by using mean, standard deviation, 25th and 75th percentiles with PCA. The proposed approach was evaluated by confusion matrix, F1 score, ROC curves, and AUC metrics further to show the robustness across different emotional classes. The results demonstrated that Wav2TP significantly outperformed many state-of-the-arts SER systems, offering a lightweight yet effective alternative.</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

Wav2TP: A novel speech emotion recognition model using temporal pooling over transformer-based Wav2Vec2 embeddings

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Yunus Korkmaz,
  • Yaser Jararweh

摘要

Speech Emotion Recognition (SER) has emerged as a crucial component in enhancing machine-user communication, particularly in conversational AI platforms and mental health monitoring systems. Despite the recent progress enabled by deep learning and pretrained speech models, existing SER systems still struggle with generalizability, temporal context preservation, and limited interpretability of acoustic features. This paper introduces Wav2TP, a novel SER framework that leverages temporal pooling techniques on contextualized embeddings extracted from the transformer-based Wav2Vec2.0 network. The proposed model incorporates statistical pooling strategies which are mean, standard deviation, 25th and 75th percentiles combined with Principal Component Analysis (PCA) to reduce redundancy and enhance the feature representation. A Multi-Layer Perceptron (MLP) classifier was subsequently used for emotion classification. The proposed method was then evaluated on the EmoDB dataset under four different configurations. The baseline model using mean pooling achieved an accuracy of 88.79%, while adding attention layers and additional pooling statistics progressively improved performance. The best-performing configuration achieved a 94.39% test accuracy by using mean, standard deviation, 25th and 75th percentiles with PCA. The proposed approach was evaluated by confusion matrix, F1 score, ROC curves, and AUC metrics further to show the robustness across different emotional classes. The results demonstrated that Wav2TP significantly outperformed many state-of-the-arts SER systems, offering a lightweight yet effective alternative.