The Efficient Music Identification for FM Broadcast Monitoring Using MFCC-based Space-saving and Robust Audio Fingerprinting

Abstract

Audio identification techniques for unknown songs in today music industry are very popular for their auto detection ability to small pieces of audio signals. The research methodologies for audio identification systems vary based on the acoustics features extraction methods such as Mel Frequency Cepstral Coefficients (MFCC), Bark scale acoustics features, Filter Bank Energy (FBE), etc. Extracted features are represented as a compact and small form of audio, in cases well known as audio fingerprints. Audio fingerprint extraction is the main technique for audio identification system which is used by large international music companies such as Gracenote, Pandora, Apple music. One of the main features of audio fingerprinting is the detection of full songs by small pieces of audio which only need to take between 3 seconds to 10 seconds according to granularity and robustness ratio. As the digital age is changing to streaming style instead of buying songs by one from online distribution platforms, digital streaming companies like YouTube has been facing issues to make sure rules and regulations for benefit sharing to contents owners. After changing the music distribution style from CD selling into streaming in digital platforms, the authors and content creators have more chances to get benefits from their own contents so-called property. Unfortunately, our country Myanmar is still in progress to make precise laws and regulations to protect artists and other content owners from those who copy the contents illegally. Myanmar is changing its music distribution style from CD selling to online music platforms since 2011, in this year, illegal copyright infringement cases were committed. Founder of Legacy Music Network Company Limited, Dr. U Ko Ko Lwin said that the distribution market is breaking down to these violations beyond ethics, and so the concerned artists get unfair benefits. Almost all of the music industry in Myanmar has changed into online music distribution style after 2015. FM broadcasting is one of the big businesses in Myanmar. Various songs are broadcast daily including old and classic songs. After the CD distribution market is changed, the audiences are more interested in streaming music and videos. For the iv audience who wants to know which songs he or she listens to is the technical challenge in audio fingerprint extraction. Therefore, the audio identification system which is used by audio fingerprinting extraction methods is needed to automatically detect songs and their related contents from broadcasting FM audios. Moreover, the Myanmar music industry urgently needs an efficient broadcast monitoring system to solve copyright infringement issues and illegal benefit-sharing between artists and broadcasting stations. In this thesis, a broadcast monitoring system is proposed for Myanmar FM radio stations by utilizing space-saving audio fingerprint extraction based on the Mel Frequency Cepstral Coefficient (MFCC). This study focused on reducing the memory requirement for fingerprint storage while preserving the robustness of the audio fingerprints to common distortions such as compression, noise addition, etc. In this system, a 3-second audio clip is represented by a 2,712-bit fingerprint block. This significantly reduces the memory requirement when compared to Philips Robust Hashing (PRH), one of the dominant audio fingerprinting methods, where a 3-second audio clip is represented by an 8,192-bit fingerprint block. The proposed system is easy to implement and achieves correct and speedy music identification even on noisy and distorted broadcast audio streams. In this research work, we deployed an audio fingerprint database of 7,094 songs and broadcast audio streams of four local FM channels in Myanmar to evaluate the performance of the proposed system. The experimental results showed that the system achieved reliable performance.