Impacto da Variedade Instrumental na Performance de Modelos de Alinhamento ÁudioPartitura

Abstract

The alignment of audio with its score is a challenging area of audio processing, especially when it involves a wide variety of instruments and styles. This study investigates the impact of instrumental variety on the training of audio-to-score alignment models using Siamese Neural Networks (SNN) and Dynamic Time Warping (DTW). The central hypothesis is that the diversity of instruments in the training dataset improves the accuracy and effici ency of the alignment. To test this hypothesis, audio databases of different instruments, including drums, guitar, piano, and flute, were used, and separate datasets were created for each type of instrument, as well as a mixed dataset. The models were trained and evalu ated in intra-instrumental, inter-instrumental, and mixed scenarios. The results indicated that intra-instrumental characteristics are more relevant for model generalization and that there is a performance degradation in diluted inter-instrumental databases, highlighting the importance of data quality within the same instrumental scope. Models trained purely on individual databases showed better performance in generalizing to other instruments and in mixed scenarios.