Rede Transformers Aplicada na Detecção de Voz Cantada

Abstract

Singing voice detection (SVD) aims to identify vocal excerpts in songs and is essential in areas of musical information retrieval (MIR) related to vo cals, such as singer identification and in pre-processing processes, being used in the separation of singing voices, automatic transcription of song lyrics, among other applications. Although singing detection seems like a simple task for hu mans, it proves to be extremely demanding for machines, with its main challen ges being the complexity of sound patterns, frequency overlap and variations in singing style. However, with the advancement of deep learning models and audio manipulation techniques, the results have improved significantly over the years. This work aims to demonstrate the impact on SVD of using Transformers in two datasets: the Jamendo Corpus and DALI. The results showed that the mo del was more difficult to train with the second, presenting a difference of about 10% in the F1-score compared to the first, which was expected due to the lower accuracy and balance of this set. Furthermore, using Demucs, a music source separation, to separate voices from instrumental accompaniment improved vo cal detection accuracy. The comparison highlights how different data volumes and qualities influence the model’s performance in the SVD task.