Spoken Digit Recognition using Machine and Deep Learning-Based Approaches

Recognition of spoken words is one of the fields that attract researchers due to its importance. It involves in various applications as many applications can be fed with data in voice format. One of vocal recognition fields is Spoken Digit Recognition (SDR) which aims to translate the input spoken audio file to its relative numerical value. Machine Learning (ML) and Deep Learning (DL) are used in various digital signal processing applications due to their ability to extract useful features from input data which enhances the accuracy of the target models as well as their ability to adapt with wide range of applications. This paper proposes two approaches to tackle the problem of spoken digits recognition. First approach is to recognize the spoken digits using wavelet time scattering and Support Vector Machine (SVM) classifier. While the second aims to solve this problem using Mel-frequency spectrograms and Deep Convolutional Neural Networks (DCNN). Experiments are performed with Free Spoken Digit Dataset (FSDD) as a training and testing dataset. Although the second approach outperforms the first one by an average increase of 1.1% in terms of accuracy metric, results shows that both approaches are suitable for such an application.