Classification of Human and Synthesized Audio in Marathi Speech using Melspectrogram Analysis and Convolutional Neural Network (CNN) Model

In recent years, the demand for effective audio classification systems has grown significantly, due to the rise of audio synthesizer tools and text-to-speech converters. Though there are many developments in detecting audio in English speeches, there is an urgent need for development in regional languages too. This research focuses on classifying synthesized and real human audio of Marathi speech using Deep Learning techniques. We used Melspectrograms-based pre-processing to extract key features such as formants and harmonics. By using Mel-spectrograms for a dataset of 28,863 samples consisting of both human and AI-synthesized speech were employed to capture temporal and perceptual frequency patterns in audio signals. We trained and evaluated a convolutional neural network (CNN) model using a split of 70% training, 15% testing and 15% validation data on these spectrograms. With an exceptional accuracy of 99.68%, our model indicated an excellent proficiency to classify between original human and synthesized audio. Our work will potentially mitigate risks associated with digital deception in diverse linguistic contexts.