Singing Evaluation based on Deep Metric Learning

This paper aims to evaluate singing performance based on deep metric learning. As the vocal sound will be the input, we will first need to separate that from a soundtrack. After the separation, the vocal audio will be represented by Mel-spectrogram as an input in our proposed model. The process to build up our model splits into pre-training and training steps. Meta learning is adopted for pre-training while deep metric learning is adopted for training. The output of the model is a Euclidean distance reflecting the singers' performance, which is determined by comparing their sounds to the originals. Experimental results show a stable and reliable singing evaluation.