Searching Music Using Lyrics Extraction

Abstract

More and more audio content is added to the Internet each and every day, but techniques available for audio searching have not been able to reach a satisfactory level. The problem is even more complicated when it comes to musical audio. Rather than depending on meta-data for music search, the world needs to come up with alternative ways to search the real content of musical audio. We propose a novel method for searching binary music files, where ordinary search engines could not access. We argue that if the singing voice of the music files can be extracted, it can be stored as text, and these stored lyrics could be used as an indirect measure to search the content of music files. We also present a system that would extract the phonemes from songs in o ne mode, store it for lates use, and search the stored data online to give results for the user queries. Thus the system has two modules: The o ine Lyrics Extractor, and the online Search Module. We also discuss how this system can operate as a collection of web services in the modern internet. We discuss the training and testing of the system using two data sets. First data set cosisits of singing voice and the second data set consists of the same lyrics in the singing voice databse narrated as normal speech. We present the test results and the observations of experiments on these two data sets. We show that firstly, for any kind of input, phoneme recognition increases the percentatge of all errors compared to speech recognition. Secondly, for any kind of recognition task, singing voice introduces more errors in every form of errors compared to normal speech. Thirdly, for any kind of input (singing voice or normal speech) phoneme recognition is harder than speech recognition. Fourthly, for any kind of recognition task (phoneme recognition or speech recognition), recognizing singing voice is harder than recognizing normal speech. We also analyze the reasond behind the phenomenons on which we came into the four conclusions above. We also give reasons for why this kind of system is not possible for certain type of audio formats. We also show that the recognition rate is lower that expected, and the reason behind this accuracy drop.