Convolutional Neural Network in Combination with Multiple Machine Learning Models to Recognize the Spoken Digits

The classification of audio signals plays a crucial role in various applications including speech recognition, analysis of environmental audio, and music classification. Digit recognition, as a subset of audio classification, plays an essential role in domains comprising automated systems, telecommunications, transport systems, and security. A review of recent research papers in the field of audio classification, focusing on the use of Artificial Intelligence such as Neural Networks, and ML - Machine Learning, along with its hybrid models, has been conducted. The objective is to enhance training accuracy using the Mel Spectrogram in CNN - Convolutional Neural Networks and improve prediction accuracy with ML models. Various hybrid approaches, including CNN combined with Logistic Regression, K-Nearest Neighbor, Random Forest, Decision Tree, Naive Bayes, Support Vector Machine, Gaussian Mixture Model, and Linear Discriminant Analysis, have been evaluated. The proposed method, CNN combined with Logistic Regression, establishes the highest accuracy compared to the other models. Experimentation was conducted using the Audio MNIST dataset, which includes 10 distinct classes of digits. Numerical results reveal that the CNN + Logistic Regression accomplished the highest Accuracy value of about 99%, and CNN + K-Nearest Neighbor achieved 98%, making CNN + Logistic Regression the most effective model for digit recognition within the audio classification field.