Abstract
Speaker Recognition (SR) systems suffer from severe performance degradation when deployed in real-world environments with ambient noise and reverberation. This study investigates a joint harmonic-percussive feature representation, combining RASTA-MFCC descriptors extracted independently from both harmonic and percussive streams that help to manage noise in classical machine learning classifiers (SVM, Random Forest, Naïve Bayes), and compares them with the deep embedding extractor ECAPA-TDNN. Evaluated on the VOiCES dataset across diverse room geometries and background distractors, our results demonstrate that while deep neural models handle adverse acoustics natively, HPPS provides a consistent performance boost for classical ML pipelines with a decrement of the error rate up to 5% on weighted mean error and 8% on specific noises as television, highlighting a viable mechanism for enhancing robustness at the expense of additional computational overhead.