Optimizing Mel-Frequency Cepstral Coefficients for Improved Robot Speech Command Recognition Accuracy

Abstract

In this study, we optimized Mel-Frequency Cepstral Coefficients (MFCC) for speech recognition. The experiment involved recording eight words by one hundred individuals with tempo, pitch, duration, resolution, and speech clarity variations. The results showed that the image's shape for each word remained consistent despite variations in pronunciation, confirming that MFCC is an effective method for feature extraction in speech recognition, particularly in robotic environments. Using logarithms in MFCC clarifies minor differences in amplitude scale, resulting in sharper contours and more visible details. This study recommends optimizing MFCC parameters, testing on larger datasets, and applying the method in various operational conditions to enhance its reliability in the future