Chromagram Features Analysis for Learning-Based Query by Humming Systems

The Query by Humming (QBH) system is a melody-based searching system that can retrieve the song without using the information of the title, the composer, or lyrics. To well distinguish the songs with similar melodies, an accurate QBH system is required. Three primary techniques dominate the performance of audio analysis: (i) onset detection and segmentation, (ii) fundamental frequency analysis, and (iii) melody matching. In this study, we mark a departure from conventional approaches by adopting the Chromagram as the primary feature, coupled with the integration of deep learning techniques to enhance the accuracy of QBH. In music analysis, the Chromagram is served as a pivotal representation of the pitch content of an audio signal over time, facilitating the breakdown of the signal into individual pitch classes. Each pitch class is typically represented as a vector element, encompassing notes such as $\mathrm{C}$, C#, and D. This representation is helpful for chord recognition, key detection, melody extraction, and music similarity comparison. The Chromagram can efficiently capture the harmonic content of a music signal and effectively improve the performance of music retrieval and analysis.