Two-Way Feature Extraction for Speech Emotion Recognition Using Deep Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Automated human emotion recognition from speech is a challenging task.
• Deep learning models offer potential but require effective feature extraction.
• Previous research often relied on single feature extraction methods.

Purpose of the Study:

• To investigate and compare two distinct feature extraction techniques for improved speech emotion recognition.
• To evaluate the efficacy of principal component analysis (PCA) versus mel-spectrograms for emotion identification.
• To assess the performance of deep neural networks (DNNs) and pre-trained VGG-16 models on extracted features.

Main Methods:

• Proposed a two-way feature extraction approach using super convergence.
• Extracted numerical features using PCA, followed by a deep neural network (DNN).
• Generated mel-spectrogram images and utilized a pre-trained VGG-16 model.

Main Results:

• Extensive experiments and comparative analysis were conducted on two datasets.
• The mel-spectrogram approach with VGG-16 demonstrated superior accuracy compared to PCA-based DNNs.
• The RAVDESS dataset yielded significantly better results with the mel-spectrogram method.

Conclusions:

• Mel-spectrogram image representation combined with pre-trained models like VGG-16 is a highly effective strategy for speech emotion recognition.
• The choice of feature extraction method significantly impacts the performance of emotion recognition systems.
• Further research can build upon these findings to develop more robust and accurate emotion detection technologies.