Scholarlink

This paper presents a method for detecting three basic emotions – happiness, fear, and anger – through the analysis of screams using a convolutional neural network (CNN). To induce these emotions, ten subjects were instructed by a certified acting professor to produce three different screams, each correlated with one of the emotions, in a soundproof room equipped with professional equipment. The CNN architecture utilizes mel-frequency cepstral coefficients (MFCCs) extracted from speech signals, with a total of 87,944 trainable parameters. For this study, from a total of 153 samples, 102 samples were used for training the CNN, and 51 samples were used for the testing part. The network then classifies the emotions into eight categories, including one neutral, three positive, and four negative classes, and also provides information on the intensity of each detected emotion. The results of the CNN testing showed recognition rates of 66.7% for happiness, 58.8% for fear, and 68.8% for anger.

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

Emotion Recognition in Screaming Audio Files with Convolutional Neural Network (CNN)

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Costin Andrei Bratan,
  • Alexandra Andrei,
  • Claudia Tocila - Matasel,
  • Bogdan Morosanu,
  • Bogdan Ionescu,
  • Ana Voichita Tebeanu,
  • Monica Dascalu,
  • Gabriela Bobes,
  • Ionut Popescu,
  • Adriana Apostol,
  • Eduard Franti,
  • Gheorghe Iana,
  • Gabriela Iorgulescu

摘要

This paper presents a method for detecting three basic emotions – happiness, fear, and anger – through the analysis of screams using a convolutional neural network (CNN). To induce these emotions, ten subjects were instructed by a certified acting professor to produce three different screams, each correlated with one of the emotions, in a soundproof room equipped with professional equipment. The CNN architecture utilizes mel-frequency cepstral coefficients (MFCCs) extracted from speech signals, with a total of 87,944 trainable parameters. For this study, from a total of 153 samples, 102 samples were used for training the CNN, and 51 samples were used for the testing part. The network then classifies the emotions into eight categories, including one neutral, three positive, and four negative classes, and also provides information on the intensity of each detected emotion. The results of the CNN testing showed recognition rates of 66.7% for happiness, 58.8% for fear, and 68.8% for anger.