A Hybrid Cepstral–Wavelet Feature Framework with Neural Network Classification for Real-Time Spoken-Digit Recognition

Abstract

Automatic speech recognition (ASR) remains challenging due to speaker variability, noise, and the complexity of speech signals. This  study presents a lightweight hybrid framework for isolated Persian spoken-digit recognition that integrates Mel-frequency cepstral  (MFCC-based) preprocessing, the discrete wavelet transform (DWT), sequential UTA (S-UTA) feature selection, and a multilayer  perceptron (MLP) classifier, implemented in LabVIEW. The approach combines spectral and time–frequency feature extraction to obtain  a compact representation of the speech signal, while feature selection removes redundant sub-band information and improves learning  efficiency. Rather than proposing a new learning architecture, the contribution of this work lies in (i) a multi-resolution front-end that  fuses silence-trimmed cepstral and delta-cepstral statistics with wavelet-packet sub-band energies, and (ii) S-UTA, a stability-corrected  sequential formulation of the UTA feature-selection algorithm with baseline re-estimation after every accepted removal, integrated into  a unified, computationally efficient pipeline that supports real-time (online) operation in resource-constrained environments. The  evaluation is conducted on the public PSDR Persian spoken-digit corpus (900 utterances: 10 digits, 9 speakers, 10 repetitions) using 5 fold cross validation. The full pipeline attains 87.3 ± 1.3% accuracy under the speaker-dependent protocol with per-class specificity above  97.8%, while S-UTA prunes 10–21% of the features at no significant accuracy cost; under identical conditions the classical one-shot UTA  collapses to 67.3 ± 19.1%, retaining  as few as 4 of 136 features on some folds. An SVM trained on the same features reaches 95.2%, and  speaker-independent accuracy is 48–55%; both findings are reported and discussed openly. The results indicate that the proposed pipeline  is a practical lightweight solution for speaker-dependent, real-time isolated-word recognition in embedded and low-resource settings.