*3.3. Feature Selection*

Feature selection is a process of filtering insignificant features to create a subset of the most discriminating features in the input data. It reduces the cause of dimensionality and improves the success of a recognition algorithm. Based on the literature review, we observed that most methods used for feature selection are computationally expensive because of the enormous processing time involved. Moreover, accuracy and precision values of some of these methods are reduced because of redundancy, less discriminating features and poor accuracy values in certain emotions such as the fear emotion. The authors have selected 133 MFCC spectral features for the MFCC1 and 90 features based on MFCC, ZCR, energy and fundamental frequency for the MFCC2. In a nutshell, the e ffectiveness of the HAF was tested against one homogenous set of acoustic features (MFCC1) and another hybrid set of acoustic features (MFCC2).

Table 2 presents the list of prosodic and spectral features of this study in terms of group, type and number of features with their statistical characteristics obtained with the aid of jAudio that generated a csv file with numerical values for each feature together with the respective emotion class. The literature inspired brute force approach was then explored to select non-redundant features that are likely to help improve emotion recognition performance. This implies inspecting through a series of experiments for possible combinations of features to discover the right features that define HAF. This approach is although laborious to require the help of a stochastic optimization technique, it has yielded excellent performance results in this study.


**Table 2.** List of the selected prosodic and spectral features.
