The Design And Implementation Of Fingerprint-based Music Humming Retrieval System

With the development of the Internet and the progress of multimedia compression, digital music starts to pop up. In this case, the music retrieval becomes mass data retrieval. Traditional retrieval method are based on the keywords marked manually. In order to retrieve songs, users have to remember the key words. This kind of method can’t satisfy the users’ demand to retrieve a song without any key words, so a method is in urgent need. Query by humming is one of the content-based retrieval. Instead of adding extra information to the music itself, the humming retrieval method based on fingerprint takes advantage of the characteristics of music, such as tone, melody and rhythm to retrieve music. In this way, just a small piece of humming music clip can help users retrieve the songs they wanted. This method provides a natural way for retrieval and it’s widely used in mobile platforms.This paper mainly focus on the research of query by humming and the music humming retrieval system based on fingerprint. The main work of this paper is described as following aspects:1) Implement the audio fingerprint extraction algorithm based on wavelet transform.The main steps of audio fingerprint extraction algorithm based on wavelet transform are as follows:firstly, convert audio signal into spectrogram; secondly, extract fingerprints by applying wavelet transform to the spectrogram, because wavelet transform can excellently denote the characteristics of image information in time domain and frequency domain; finally, produce leaner fingerprints by dimension reduction algorithm. This algorithm is applied to query by humming system for the first time. When quality of input humming clip is good enough, the retrieval rate can be more than 90%.2) Implement and improve audio fingerprint retrieval and matching algorithm based on local sensitive hash (LSH)Audio fingerprint retrieval algorithm based on LSH is the most popular high dimensional rapid retrieval algorithm so far. In this paper, we added a step named second voting in the process of LSH algorithm in order to increase the number of candidate sequences and improve the retrieval accuracy. When the parameters are suitable, the retrieval rate is higher than the unimproved system.3) Design and implement a humming retrieval system.Humming retrieval system can be divided into two parts, one part is establishing music fingerprint database, the other is retrieving the humming fragments. According to the overall process of system, the function of system can be divided into four parts which are audio inputting, retrieving, data building and returning. MVC pattern is used as the overall architecture of this system. In this paper, according to the design of the system, combined with fingerprint extraction algorithm and audio fingerprint retrieval and matching algorithm, we implemented a query by humming system that can record, upload and retrieve humming clips efficiently.