A Music File Detection Method based on Convolutional Neural Network for Video-on-Demand Platform

On the video-on-demand (VOD) platform, audio processing includes noise suppression, loudness normalization, codec and others. However, these processing are not suitable for all scenes, especially for music audio quality and subjective perceptual. Therefore, different processing will be selected according to the music detector result. In this paper, a music or non-music audio detection used convolution neural network based on the Mel-spectrogram and modified spectral flux with music frame proportion estimation method is proposed. In the method, the Mel-spectrogram flux will greatly improve the traditional spectral flux calculation. This paper creatively proposes a music file-wise detection algorithm based on the VOD platform compared with previous frame-wise detection method. Proposed method has strong robustness, high computing efficiency, and low delay. It is more suitable for VOD platform audio detection requirements, and has direct benefits for audio quality control.