Scholarlink

<p>Speech Emotion Recognition (SER) has gained more significance in<!--Query ID="Q1" Text="Please check if the section headings are assigned to appropriate levels. " Resolved="yes"--> the field of Human-Computer Interaction. Many studies have been conducted so far, but performance remains a challenge due to the complexity of natural language and the features involved. Emotion recognition systems have a low level of accuracy, because it is difficult task to extract features from speech emotion databases. In this work, the accuracy of SER system is improved by the use of robust hybrid feature extraction methods and classification models. To identify the emotions in speech utterances, SER uses a<!--Query ID="Q2" Text="Please check if the author names and affiliations are captured and presented correctly. " Resolved="yes"--> two-stage process. First Feature Extraction, followed by Classification Model. Hybrid feature extraction with fusion is used in the process of extracting features. Here 39 MFCC’s (13 Mel Frequency Cepstral Coefficients (MFCC), 13 ∆ MFCC, 13 ∆∆ MFCC are included as acoustic features and Deep Features also extracted from Pre-trained Convolution Neural Network. Then Auto-Encoder approach is used to choose suitable features from previously extracted features. Support Vector Machines are used in the second step of the classification model (SVM). The research work uses the RAVDESS- Ryerson Audio-Visual Database of Emotional Speech and Song database from the Ryerson Multimedia Laboratory as well as the native specially created database called DETL-Database for Emotions in Telugu Language. We achieved 52.11% accuracy using MFCC features, 70.73% with hybrid features (MFCC + ΔMFCC + ΔΔMFCC), and 81.23% with deep acoustic features on the RAVDESS database. For the DETL database, deep acoustic features yielded an accuracy of 82.37%.</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

Hybrid acoustic-deep features with auto encoders for speech emotion recognition

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Kogila Raghu,
  • Manchala Sadanandam,
  • Bh Hanumanthu

摘要

Speech Emotion Recognition (SER) has gained more significance in the field of Human-Computer Interaction. Many studies have been conducted so far, but performance remains a challenge due to the complexity of natural language and the features involved. Emotion recognition systems have a low level of accuracy, because it is difficult task to extract features from speech emotion databases. In this work, the accuracy of SER system is improved by the use of robust hybrid feature extraction methods and classification models. To identify the emotions in speech utterances, SER uses a two-stage process. First Feature Extraction, followed by Classification Model. Hybrid feature extraction with fusion is used in the process of extracting features. Here 39 MFCC’s (13 Mel Frequency Cepstral Coefficients (MFCC), 13 ∆ MFCC, 13 ∆∆ MFCC are included as acoustic features and Deep Features also extracted from Pre-trained Convolution Neural Network. Then Auto-Encoder approach is used to choose suitable features from previously extracted features. Support Vector Machines are used in the second step of the classification model (SVM). The research work uses the RAVDESS- Ryerson Audio-Visual Database of Emotional Speech and Song database from the Ryerson Multimedia Laboratory as well as the native specially created database called DETL-Database for Emotions in Telugu Language. We achieved 52.11% accuracy using MFCC features, 70.73% with hybrid features (MFCC + ΔMFCC + ΔΔMFCC), and 81.23% with deep acoustic features on the RAVDESS database. For the DETL database, deep acoustic features yielded an accuracy of 82.37%.