Impact of Feature Normalization Techniques for Recognition of Speech for Mathematical Expression

Speech recognition is an important aspect of Human–Computer Interaction (HCI). Speech recognition systems can be implemented with the help of speech signal patterns. In pattern recognition, feature selection and feature extraction are important phases. The performance of the model depends not only on the feature selection and feature extraction but also on feature preprocessing. In our existing work, we have proposed a hybrid feature extraction algorithm for the recognition of speech for audio mathematical expression. In this paper, we have studied five feature normalization techniques used for audio processing. We compared five feature normalization functions with time domain, frequency domain, and cepstral domain features using Audio Mathematical Expression Recognition (AMER) dataset. Six types of audio mathematical expression are classified using five feature normalization functions. The observations indicate that the standard scaler and robust scaler give more accuracy and quantile transformer scaler gives less accuracy. The combination of feature normalization, feature extraction, and classification complies with higher accuracies.