Evaluating Interval-based Tokenization for Pitch Representation in Symbolic Music Analysis
Proposed interval-based tokenization for symbolic music, improving model performance and interpretability.
Key Findings
Methodology
This study introduces an interval-based tokenization strategy that combines absolute pitch and interval information to better capture melodic contour and harmonic relations. Evaluations on three music analysis tasks demonstrate its effectiveness.
Key Results
- Interval tokenization improved performance by 6% in melody start detection.
- In chord inversion identification, bottom-line reference interval tokenization performed best.
- In era classification, interval tokenization increased accuracy by 1.2%.
Significance
The study addresses the limitations of traditional absolute pitch tokenization by introducing interval-based tokenization, enhancing model performance and interpretability in music analysis, advancing the field of symbolic music analysis.
Technical Contribution
Introduced a general framework for interval-based tokenization, applicable to various music analysis tasks, offering new perspectives for musical interpretation.
Novelty
First to apply interval tokenization to polyphonic music symbolic representation, overcoming limitations of traditional absolute pitch tokenization, providing a representation more aligned with human musical perception.
Limitations
- Interval tokenization's application in generative tasks is limited, especially horizontal interval tokenization.
- Experimental framework is simple, not fully leveraging interval tokenization strategy.
Future Work
Future work could explore more combinations of interval tokenization strategies, especially in generative tasks, and integration with other time encoding strategies.
AI Executive Summary
Symbolic music analysis often employs models from natural language processing, such as Transformers, which require input data to be represented as sequences. Traditional tokenization strategies rely on absolute MIDI values to represent pitch information, potentially overlooking relationships between notes. This study proposes an interval-based tokenization strategy, demonstrating through evaluations on three music analysis tasks that this approach can enhance model performance and interpretability.
The research shows that interval tokenization excels in melody start detection, chord inversion identification, and era classification tasks, particularly improving performance by 6% in melody start detection. This indicates that interval tokenization better captures melodic contour and harmonic relations, offering new perspectives for musical interpretation.
While interval tokenization performs well in analysis tasks, its application in generative tasks remains limited, especially horizontal interval tokenization. Future research could explore more combinations of interval tokenization strategies, particularly in generative tasks, and integration with other time encoding strategies to further enhance the performance and application scope of symbolic music analysis.
Deep Analysis
Background
Symbolic music analysis is a crucial research area in music information retrieval, often utilizing sequence models from natural language processing, such as Transformers. These models require music data to be represented as sequences, and traditional tokenization strategies mainly rely on absolute pitch encoding, which may overlook relationships between notes.
Core Problem
Traditional absolute pitch tokenization methods may overlook relationships between notes, failing to fully capture melodic contour and harmonic relations, thus limiting model performance and interpretability.
Innovation
This study proposes an interval-based tokenization strategy, combining absolute pitch and interval information to provide a representation more aligned with human musical perception, overcoming limitations of traditional absolute pitch tokenization.
Methodology
- �� Introduced a general framework for interval-based tokenization, combining absolute pitch and interval information.
- �� Evaluated interval tokenization on melody start detection, chord inversion identification, and era classification tasks.
- �� Used MidiTok package for initial tokenization and conducted experiments on BERT models.
Experiments
Experiments were conducted on multiple datasets, including POP909, MTC-Piano, and OpenScore String quartets. BERT models were used to evaluate the performance of interval tokenization in melody start detection, chord inversion identification, and era classification tasks.
Results
Interval tokenization improved performance by 6% in melody start detection. In chord inversion identification, bottom-line reference interval tokenization performed best. In era classification, interval tokenization increased accuracy by 1.2%.
Applications
Interval tokenization can be applied to various music analysis tasks, such as melody start detection, chord inversion identification, and era classification, enhancing model performance and interpretability.
Limitations & Outlook
Interval tokenization's application in generative tasks is limited, especially horizontal interval tokenization. The experimental framework is simple, not fully leveraging the interval tokenization strategy.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. Traditional absolute pitch tokenization is like focusing only on the exact weight of each ingredient, ignoring how they pair together. Interval tokenization is like paying attention to how flavors combine, helping you better understand the overall taste of the dish.
ELI14 Explained like you're 14
Imagine you're playing a music game. Traditional tokenization is like only looking at the height of each note, ignoring the distance between them. Interval tokenization is like focusing on the distance between notes, helping you better understand the melody and harmony. Isn't that cool?
Glossary
Tokenization
The process of representing input data as sequences.
Used to convert music data into sequences that models can process.
MIDI
A digital interface for representing musical notes.
Traditional tokenization strategies rely on absolute MIDI values.
Interval
The relative distance between notes.
Used to represent melodic contour and harmonic relations in music.
Transformer
A deep learning model for processing sequence data.
Used for symbolic music analysis tasks.
BERT
A pre-trained language model based on Transformer.
Used to evaluate the performance of interval tokenization.
Open Questions Unanswered questions from this research
- 1 Interval tokenization's application in generative tasks remains limited, especially horizontal interval tokenization.
- 2 How to combine interval tokenization with other time encoding strategies to enhance performance?
Applications
Immediate Applications
Melody Start Detection
Using interval tokenization to enhance the accuracy of melody start detection.
Long-term Vision
Music Generation
Exploring the application of interval tokenization in music generation tasks to improve the quality of generated music.
Abstract
Symbolic music analysis tasks are often performed by models originally developed for Natural Language Processing, such as Transformers. Such models require the input data to be represented as sequences, which is achieved through a process of tokenization. Tokenization strategies for symbolic music often rely on absolute MIDI values to represent pitch information. However, music research largely promotes the benefit of higher-level representations such as melodic contour and harmonic relations for which pitch intervals turn out to be more expressive than absolute pitches. In this work, we introduce a general framework for building interval-based tokenizations. By evaluating these tokenizations on three music analysis tasks, we show that such interval-based tokenizations improve model performances and facilitate their explainability.