Improving Query by Humming System using Frequency-Temporal Attention Network and Partial Query Matching

Search engine technology makes it easy to find information from the many available sources. One of the available information is music. In Music Information Retrieval, Query by Humming is an effective and natural method for searching music in databases. The Unified Algorithm is the latest research in this field. The QbH system is divided into two stages: melody extraction and matching. Several studies have found that the statistical approach used in the Unified Algorithm did not perform better than the data-driven approach. In addition, the song that the user hums often is only a part of the whole melody. This causes a match to be made between the humming query and the query containing the entire melody. This study proposes a modification to the QbH system by using a data-based approach for melody extraction and matching only a part of the melody. The latest research on the topic of data-based melody extraction is the Frequency-Temporal Attention Network (FTANet). The combination of FTANet as melody extraction and modification of the unified algorithm can provide better performance compared to the baseline system. However, in terms of computational time, both melody extraction and melody matching processes take much longer. This study also revealed that the Unified algorithm is not suitable for use in systems with large datasets.