Leveraging Ensemble and Deep Learning Techniques for Spoken Odia Digit Recognition

Spoken digits recognition (SDR) is a supervised automatic speech recognition method widely used in human-machine interaction applications, including dialing systems, financial transactions, airline reservations, and price extraction tasks. However, designing SDR is a challenging endeavour that involves the collection of labelled audio data, the appropriate selection of feature extraction techniques, and the creation of the top-performing model, especially for Indian languages like Odia.This paper examines the integration of ensemble machine learning, deep learning, and hybrid ensemble approaches to enhance spoken Odia digit recognition. Initially, we started by compiling a dataset of isolated voiced Odia digits. Subsequently, this dataset underwent preprocessing, which included normalization, noise reduction, and silence removal. In order to retrieve the mean and variance as feature vectors, a variety of feature extraction methods are employed for each spoken digit file. Our research examines the efficacy of several machine learning models in conjunction with ensemble methods, including bagging, boosting, stacking, and voting to aggregate the basic machine learning models’ predictions. We also explored the hybrid ensemble method, which uses 1-D CNN in conjunction with several machine-learning models. According to the experimental findings, ensemble approaches perform noticeably better than individual models in terms of generalization, accuracy, and dependability. Specifically, hybrid ensemble approaches yield the highest improvement by combining multiple model predictions to achieve superior performance with an accuracy of 91% for classifying spoken Odia digits.