Digital Preservation and Noise Reduction using Machine Learning

This paper proposes a digital preservation solution for Sinhala audios to conserve those as documents with noise reduction. The solution has implemented multiple noise reduction techniques as a pre-processing step to remove unwanted internal and external noises. A two-step, two-way noise reduction process is applied to produce clean audios based on Deep Convolutional Neural Network (DCNN) and adaptive filter-based techniques. This approach implements two separate noise reduction models for internal and external noises. After that, the speech recognition decoder recognizes the speech and converts it to a Unicode document by acoustic, language, and pronunciation models using extracted audio features from the denoised audio. Further, noise reduction models are decoupled from the preservation solution and exposed as a sub solution for multilingualism noise reduction, supporting English and Sinhala audios.