RETRACTED ARTICLE: Impact of autoencoder based compact representation on emotion detection from audio

Area of Science:

• Speech processing
• Machine learning
• Affective computing

Background:

• Speech emotion recognition (SER) is vital for human-computer interaction.
• Existing SER systems often struggle with real-time application constraints.
• Need for efficient and accurate methods for audio feature representation.

Purpose of the Study:

• To propose and evaluate a compact audio representation using autoencoders for SER.
• To assess the impact of dimensionality reduction on classification accuracy.
• To develop systems suitable for real-time emotion recognition from speech.

Main Methods:

• Utilized conventional autoencoders for audio dimensionality reduction.
• Tested the approach on Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and Toronto Emotional Speech Set (TESS) datasets.
• Implemented Support Vector Machines (SVM), Decision Tree, Convolutional Neural Networks (CNN), Alexnet, and Resnet50 for classification.

Main Results:

• Autoencoders significantly improved classification accuracy for emotion recognition in audio.
• Compact audio representations led to efficient systems with low computing costs.
• The approach demonstrated potential for real-time SER applications.

Conclusions:

• Dimensionality reduction via autoencoders positively impacts speech emotion recognition.
• Optimizing audio feature representation is crucial for advancing SER models.
• This method offers a pathway to more practical, real-time emotion recognition systems.