Improved Feature Extraction for Sound Recognition Using Combined Constant-Q Transform (CQT) and Mel Spectrogram for CNN Input

This paper delves into an innovative paradigm for enhancing feature extraction in the realm of speech recognition. It integrates the Constant-Q Transform (CQT) and Mel Spectrogram techniques to elevate the performance of Convolutional Neural Networks (CNN) in the domain of speech recognition. The study introduces a novel augmentation to the traditional Mel-frequency cepstral coefficients (MFCC) extraction methodology. This enhancement involves utilizing the frequency representations obtained from CQT and Mel Spectrogram as inputs for the CNN model. Extensive experimentation and rigorous evaluation validate the efficacy of the proposed approach. It is evident from the results that this approach yields more intricate and meaningful spectral features when compared to the conventional MFCC technique. These findings underscore the significant potential inherent in amalgamating CQT and Mel Spectrogram techniques, showcasing their ability to significantly augment the accuracy and overall performance of CNN models in the context of speech recognition tasks. This breakthrough underscores the continuous evolution of speech recognition methodologies and presents a promising avenue for further research and development in the field.