Abstract:
In the present world, content censorship is an important concept. When it
comes to the Sinhala language, several studies have been conducted on textbased
content censorship methods, but not for audio content. The Sri Lankan
government prohibits the use of profanity in public media. Therefore, Sri
Lankan media companies must check their videos for profanity before
telecasting. Till now, this process has been done manually, and it is extremely
difficult with long videos and audio clips. This study suggests developing a
deep learning model that can automatically find profanity words in Sinhala
audio files. The ten profanity words were selected and audio samples from 100
people were gathered. The data was preprocessed, transformed into
spectrogram images, and applied to a convolutional neural network (CNN) to
develop the profanity filter model. By converting audio files to spectrograms
and applying image processing to extract the features from the dataset, the
model predicts the profanity words. This paper addresses the procedure of the
mentioned process and its capabilities with upcoming updated versions of the
final product.