TLDR: A new research system proposes a method for detecting music plagiarism in real-world audio by transcribing raw audio into musically meaningful segments. It combines various Music Information Retrieval (MIR) technologies to extract features like melody, chords, and rhythm from these segments, then calculates similarity scores based on multiple musical aspects. This segment-based approach allows for detailed analysis, partial matching, and better interpretability, addressing limitations of traditional whole-song comparison methods and offering a practical solution for music copyright protection.
In an era where music creation and distribution are more accessible than ever, thanks to continuous advancements in Music Information Retrieval (MIR) technology, the protection of music intellectual property has become increasingly vital. Addressing this critical need, a new research paper introduces an innovative system designed to detect music plagiarism in real-world audio. This system combines various MIR technologies to offer a more nuanced and effective approach to safeguarding musical copyrights.
Traditional methods for plagiarism detection often face significant challenges when applied to commercial music, which typically exists as raw audio rather than structured formats like MusicXML or MIDI. Many existing studies also tend to focus on melodically similar music or perform whole-song comparisons, which don’t always align with the complexities of real-world plagiarism cases. These cases can involve vocals, vary in length, or contain only brief plagiarized segments within much longer tracks. The proposed system tackles these issues by transcribing raw audio into musically meaningful segments, allowing for a more precise and detailed analysis.
How the System Works: Music Segment Transcription
The core of this new approach lies in its music segment transcription system. Instead of analyzing an entire song at once, the system first processes raw audio to extract essential musical components. This involves several steps:
- Source Separation: Isolating different instrumental and vocal tracks.
- Beat-Tracking and Downbeat Identification: Determining the tempo and rhythmic structure.
- Chord Recognition: Identifying the harmonic progression.
- Music Structure Analysis: Pinpointing structural changes like verses, choruses, and bridges.
Once these components are identified, the system quantizes the music into distinct segments, typically defined as four-bar intervals. Each segment is enriched with detailed musical information, including melody, chords, instruments, pitch, onset time, duration, and velocity, all mapped to specific bar numbers and positions within the score. This structured data allows for a granular understanding of the music’s composition.
Detecting Plagiarism Through Segment Similarity
With the music broken down into these information-rich segments, the system then computes similarity scores between segments from different musical pieces. This goes beyond simple melodic comparison by incorporating multiple musical aspects:
- Pattern Similarity: Based on chromagrams, which represent the intensity of pitches over time.
- Musical Complexity: A weighted count of used pitches, designed to prevent overly simple similarities (e.g., in rap music) from skewing results.
- Rhythmic Correlation: Comparing the quantized onset timings of notes.
- BPM Difference Ratio: Accounting for tempo variations between pieces.
- Chord Similarity: A composite metric that considers both Roman numeral similarity (functional harmony) and chord quality (major/minor, seventh chords).
By combining these metrics, the system can evaluate similarities comprehensively, identifying not just overall resemblances but also specific elements like a vocal melody in a chorus segment being plagiarized from another song’s chorus. For more technical details, you can refer to the original research paper: REAL-WORLD MUSIC PLAGIARISM DETECTION WITH MUSIC SEGMENT TRANSCRIPTION SYSTEM.
Also Read:
- AnalysisGNN: A Unified Framework for Comprehensive Music Score Analysis
- Deepfake Audio Detection Under Siege: A Look at Adversarial Attacks
Real-World Evaluation and Future Directions
To validate its effectiveness, the researchers compiled a “Similar Music Pair (SMP)” dataset, comprising 70 pairs of original and comparison music pieces, including known plagiarism cases and works with acknowledged influence. Experiments were also conducted using the Covers80 dataset. The evaluation involved both segment-level and song-level analyses, measuring metrics like Precision@K for segment detection and Top Average Index, Top-1 Accuracy, and Top-5 Accuracy for song-level detection.
The results demonstrate promising performance, particularly highlighting the advantages of the segment-based approach. This method provides more detailed analysis by pinpointing specific musical passages contributing to plagiarism, enables partial matching where only certain segments are similar, and offers better interpretability by showing which musical elements drive the similarity scores. This makes it particularly valuable for real-world copyright dispute applications.
While the system shows significant potential, the authors acknowledge areas for future improvement. These include enhancing individual MIR components (source separation, transcription, beat tracking) for more stable segment transcription, developing an end-to-end approach for segment transcription, and further refining the connection between segment-level and music unit-level metrics. The structured segment data format developed in this work also holds promise for other MIR tasks, such as cover song detection and music generation, paving the way for advanced music copyright protection and analysis.
