Performance Analysis of CNN-Based Spectrogram with Multiple Audio Feature Types for English Digit Recognition

Audio feature selection and neural network architecture play crucial roles in speech recognition performance. This paper presents a comparative analysis of Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs) for English digit recognition using five distinct audio feature extraction techniques: MEL, MFCC, SPECTRAL CONTRAST, CHROMA, and TONNETZ. The study utilizes the Google Speech Command dataset, with 500 samples for each of the 10 digits (0-9). We evaluate training efficiency through validation loss progression and model performance through test accuracy. Results demonstrate that CNNs processing spectrogram images consistently outperform ANNs processing feature vectors, with an average improvement of 10.87% across feature types. Among features, MFCC and MEL representations yield superior performance in both architectures, with MFCC-CNN achieving the highest overall accuracy (63.07%) followed by MEL-CNN <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(60.67 \%)$</tex>. Visual analysis of spectrograms reveals a strong correlation between the richness of time-frequency patterns and classification performance. While CNNs show significant advantages in leveraging spectro-temporal patterns, they require substantially longer training times than ANNs, presenting an important performance-efficiency tradeoff. Our findings highlight the critical interplay between feature representation and model architecture selection, providing valuable insights for audio classification system design.