Audio de-noising and quality assessment for various noises in lecture videos

Online teaching has taken up its importance post-pandemic period. Today, online teaching is considered to be one of the teaching pedagogy. This means every teacher and professor is generating online lecture videos and sharing them for students’ later use. Mostly, the environment for the video creation is in real time either in the live classroom or at home, various environmental noises interfere with the actual speech of the presenter. Therefore, there is a need for identifying the various noises that may be part of the lecture video to assess the quality of the video. Towards this, very few research works are observed. Researchers have worked on additive noises, but identifying convolutional noises is a challenge. We propose to work on the audio signal of the video lectures to identify the positions and durations of various convolutional noises and measure the amount of noise present in the audio part of the video lectures. We used various filters for identifying simultaneous talks, long silences, baby crying, kitchen sounds, and vehicle noises. The average accuracy of the proposed solution in identifying the noises and the noise positions is 97.37%. The MSE of the noise in the audio of each clip varies depending on the various noises present. This defines the quality of the audio in the lecture video.