Voice-assisted systems have become an integral component of modern automotive applications, enabling intuitive and hands-free human–machine interaction. Keyword Spotting (KWS), a fundamental task in Automatic Speech Recognition (ASR), detects predefined keywords from continuous speech and serves as the front-end for many voice-controlled systems. The quality of the extracted speech features directly influences the performance of deep learning models. This paper presents a detailed analysis of five widely used speech feature extraction techniques for speech-based keyword spotting: Log-Mel Spectrogram, Mel Frequency Cepstral Coefficients (MFCCs), Gammatone Frequency Cepstral Coefficients (GFCCs), Power-Normalized Cepstral Coefficients (PNCCs), and Perceptual Linear Prediction Coefficients (PLPCs). These feature representations are evaluated using Convolutional Neural Network (CNN) and Depthwise Separable Convolutional Neural Network (DSCNN) architectures. The models are compared in terms of classification accuracy, precision, recall, specificity, F1-score, and floating-point operations (FLOPS). Experimental results demonstrate that both 4-layer DSCNN with MFCC features and 3-layer CNN with GFCC features achieve the best trade-off between classification performance and model complexity.
 

Keyword Spotting,Feature Extraction,Log-Mel Spectrogram,MFCCs,GFCCs,,PNCCs,,PLPCs,Neural Model,,CNN,DSCNN.