Research On The Key Technology Of Query By Humming

Audio retrieval, image retrieval and video retrieval are the hot topics in the content-based retrieval fields. But the research results of the multimedia information retrieval fields is mainly focusing on image and video aspects, there are little fruits in the music retrieval fields. Now there are more and more numeric music, so the study on the music retrieval system is urgent and it can be used in many fields, such as the search engine, picking song system of KTV room, music library retrieval system and et al. The technology of query by humming can make retrieval easy to use. But there are some difficulties in the query by humming fields, such as feature extraction, selection, music melody matching, environment noise and et al.Surrounding the key technology of the query by humming, the main works in this paper are as follows:1) The system frame of the content-based music retrieval is given. After discussing the feasibility and the difficulty of the music retrieval based on contents, solution and processing frame are obtained.2) A feature extraction algorithm based on the music melody outline is proposed. The algorithm picks up the music pitch sequence, then converted it to melody outline sequence. Finally the sequence is converted to melody outline characteristics based on the table of standard tone differences value.3) A match algorithm of music retrieval is proposed. The algorithm redefines the string's similarity according to the song characteristic, and compares the string sequence to music character sequences in the database to get the object which has high similarity.4) Studying of the music database. The music database establishment method has been introduced firstly, secondly the time series method has been used in the music database, and the PAA algorithm which is a time series technology is used to optimize the music database. 5) A content-based query by humming system is designed and completed. The experimental results show that the system is more exactly in the query by music segment and the query by humming, the successful rate is over 90%. Simultaneously it has been obtained that this system does not need training process and has good robustness to environment's noise, and has higher retrieval precision.This research work is a part of the "research on the key technology of audio recognition and retrieval based on the web (No. 60673100)" supported by the National Natural Science Foundation of China.