Neural vocoder feature estimation for singing voice extraction

Abstract

Singing voice separation is the task of separating a singing voice from music . Recently, a method of generating singing voice masks using a deep learning model has been widely used, but there is a limitation in that the size of the data to be predicted is large and the reusability is poor because the spatial effect is not separated. To solve this problem, a singing voice separation method using the world vocoder was proposed, but there was a limit that the sound quality could not exceed the quality generated by the world vocoder. In this paper, we propose a singing voice separation method using a neural vocoder. A neural vocoder is a deep learning model that synthesizes voices from small-dimensional data and can generate a higher-quality voice than the world vocoder. We used a neural vocoder that takes a mel spectrogram as input. Using this, we were able to divide the singing voice separation process into a generation part and a separation part. In addition, we propose two learning methods using a voice presence. The first method is to improve the performance of the singing voice separation model by combining the voice classification model. The second method is to use voice presence data for training data so that the model can learn the characteristics of the singing voice. It was confirmed through objective evaluation that both methods were effective methods. In addition, we confirmed that our system performed higher in the objective evaluation than the singing voice separation system using the world vocoder.