QUERY BY HUMMING MUSIC INFORMATION RETRIEVAL USING DNN-LSTM BASED MELODY EXTRACTION AND NOISE FILTRATION

• Main Author: Novian Dwi Triastanto, Andreas
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/54338
• Institution: Institut Teknologi Bandung

Search engine technology has become a daily necessity. In terms of music search, the most effective and natural way to perform music search is to hum the song or so-called query by humming. In an experiment to improve the performance of a query by humming system, there are two things that can be done, namely using better melody extraction technique and noise filtration. Better melody extraction technique is needed so that the representation of the melody obtained from the query reflects more on the change of pitch while noise filtration is needed so that this system can work well in the real world where the queries tend to have natural noise. This research will try to implement the technique of DNN-LSTM based melody extraction and noise filtration with Fourier series decomposition and spectral subtraction in query by humming system. The system built will be compared with a query by humming system that uses melody extraction with PRAAT and without noise filtration. The results of this study indicate that the performance of the system that uses melody extraction with PRAAT and without noise filtration is still better in terms of the mean reciprocal rank value, top 1/3/5/10 hit ratio, and the required processing time.