Automatic Identification of Vietnamese Singer Voices Using Deep Learning and Data Augmentation

Singer identification is a fascinating and complex field in audio processing and voice recognition. It aims to determine the identity of a singer from a particular audio clip or song. This can be useful in various applications, from constructing personalised playlists to categorising and sorting large music libraries. Traditional methods of voice identification often rely on handcrafted features and statistical models. However, deep learning models, particularly neural networks, have demonstrated exceptional capabilities in automatically learning relevant features directly from raw audio data. This allows them to uncover subtle patterns and representations that were previously challenging to detect. In this paper, we present our study of the automatic identification of some Vietnamese singer voices using deep learning and data augmentation. We built a dataset consisting of ten Vietnamese artists with 2200 excerpts, then performed the identification of these artist’s voices using GRU, LSTM, and CNN models. Our research showed that the GRU model has a higher identification accuracy than the LSTM and CNN models. We proposed a new method of data augmentation for singer voice identification. This new method proved to be much more effective than the regular methods for audio data, such as noise addition and pitch shifting.