A Comparative Study of Acoustic Features for Spoken Digit Recognition in Algerian Arabic Dialects and Accented French

Abstract

This paper presents a comparative study of six acoustic feature representations for spoken digit recognition in Algerian Arabic dialects and accented French. The evaluated features include MFCC, MFCC with temporal derivatives, Log-Mel filterbank, LFCC, GFCC, and PNCC. Experiments are conducted on a corpus of isolated digits (0--99) collected from four Algerian regions (Algiers, Blida, Tizi-Ouzou, Sétif), covering both native dialectal speech and French-accented variants. Three classifiers---SVM, CNN, and BiLSTM---are compared to assess feature--model interactions. Results demonstrate that MFCC provides strong baselines across all regions, while GFCC achieves the highest accuracy (up to 94.4%) for dialectal speech. BiLSTM exhibits training instability with conventional features but improves significantly with perceptually motivated representations. These findings provide the first systematic benchmark for 100-class Algerian Arabic digit recognition and highlight the importance of feature selection in low-resource speech processing.