MazurkaBL Dataset: Expressive Piano Analysis

• MazurkaBL dataset is a comprehensive corpus of 44 Chopin Mazurkas offering high-quality beat, downbeat, and dynamic annotations.
• Its 60-second overlapping audio segments facilitate multi-scale analysis by providing extended temporal context for rhythmic and expressive features.
• The dataset underpins efficient multi-task learning models, achieving state-of-the-art performance in beat tracking, dynamic estimation, and change point detection.

The MazurkaBL dataset is the largest publicly available corpus of score-aligned solo piano recordings with both beat annotations and verified dynamic markings, specifically consisting of 44 Chopin Mazurkas. It is foundational in computational music analysis for the development and benchmarking of algorithms that require reliable ground-truth for both rhythmic (beat and downbeat) and expressive (dynamic) musical parameters.

1. Corpus Properties and Ground-Truth Annotations

MazurkaBL is distinguished by its high-quality score alignment and comprehensive annotation scheme. Each recording in the dataset features:

• Beat and Downbeat Annotations: Time-stamped beat locations allow precise evaluation of beat and downbeat tracking algorithms.
• Verified Dynamic Markings: Ground-truth dynamic levels (including pp, p, mf, f, ff) enable quantitative analysis of expressive variation.
• Solo Piano Recordings: Homogeneous instrumental context mitigates the confounding factors present in ensembles.

The dataset’s annotation fidelity is essential for multi-task predictions where rhythmic and expressive output must be jointly modeled and evaluated. The recordings serve as a benchmark for both temporal accuracy and dynamic detail.

2. Data Segmentation and Audio Processing Protocols

For methodological rigor, MazurkaBL recordings are segmented into 60-second audio clips with 50% overlap during training processes. This segmentation protocol diverges from typical beat tracking tasks, which conventionally use much shorter sequential input lengths.

Long segment duration is crucial for:

• Temporal Context: Enabling algorithms to leverage extended musical context, especially for analyses where dynamic changes relate to broader musical structure.
• Resource Efficiency: When paired with compact feature representations, longer inputs are computationally tractable, allowing large-scale analysis.

Segmentation length thus directly informs the network’s receptive field and influences performance, particularly for expressive tasks.

3. Applicability in Multi-Task Learning Architectures

The MazurkaBL dataset provides the bedrock for architectures targeting joint extraction of dynamic levels, change points, beats, and downbeats. The referenced method deploys a multi-scale network backed by Bark-scale specific loudness (BSSL) features. This approach yields:

• Parameter Reduction: With 22 Bark bands used in BSSL versus 128 Mel bins in log-Mel spectrograms, model size is reduced from 14.7M to 0.5M parameters.
• Long-Sequence Capability: Compact input enables efficient processing of lengthy audio fragments, enhancing the model’s ability to capture long-term expressive dependencies.

These properties allow simultaneous optimization for multiple outputs—dynamic level curves, change points snapped to beats, beat and downbeat locations—mapping directly to the annotated elements in MazurkaBL.

4. Evaluation Metrics and Benchmarking Framework

Algorithmic performance on MazurkaBL is rigorously quantified via F1 scores across all tasks. Evaluation protocols include:

Task	Ground-Truth Type	Metric
Dynamic Estimation	Discretized at beats	Macro F1 score
Beat/Downbeat Tracking	Beat times	F1, ±70 ms tolerance
Change Point Detection	Snapped to nearest beat	Standard F1 score

For dynamic estimation, continuous model predictions are sampled at annotated beat locations and discretized. Beat/downbeat detection requires predictions to fall within a ±70 ms window of annotated events. Change point detection aligns candidate frames to the nearest beat, conforming to the annotation paradigm.

5. Model Efficiency and Analytical Value

The MazurkaBL dataset's comprehensive annotation enables models—such as the multi-task multi-scale architecture with MMoE decoding—to be empirically validated for joint prediction efficiency. Key findings from application include:

• State-of-the-Art Performance: Models have demonstrated leading results on dynamic and change point prediction, and competitive scores on beat/downbeat tracking.
• Efficient Large-Scale Analysis: Feature design leveraging BSSL and architectural compactness (0.5M parameters) allows for scalable, resource-efficient paper of expressive performance over extended segments.
• Temporal Dependency Modeling: The dataset’s structure supports representation learning that remains sensitive to dynamics spanning entire musical phrases.

A plausible implication is that such efficiency could facilitate integration into automatic piano transcription pipelines, providing fully annotated scores from audio with minimal computational overhead.

6. Implications for Expressive Performance Research

MazurkaBL’s precise and exhaustive annotations serve as a foundation for advanced research in expressive performance analysis. Notable implications include:

• Automated Annotation of Archives: Algorithms trained and validated on MazurkaBL can be used to annotate large databases of (especially historical) solo piano recordings with expressive and metrical information.
• End-to-End System Integration: Lightweight models tailored for longer input segments are readily deployable within broader systems, such as those generating complete expressive scores from raw audio.
• Enhanced Study of Dynamics: The dataset’s inclusion of verified dynamic curves encourages rigorous investigation into the interplay between expressive loudness variation and metrical structure.

This suggests that MazurkaBL may catalyze new directions in both algorithmic development and empirical paper, particularly for understanding nuanced expressive behaviors in solo piano performance.

7. Future Directions and Research Opportunities

The MazurkaBL dataset, in conjunction with compact and efficient multi-task learning frameworks, opens several avenues:

• Scaling Expressive Analysis Beyond Chopin Mazurkas: The methodology may be extended to other composers or genres, contingent on availability of score-aligned recordings with comparable annotations.
• Expressive Rendering and Generation: Accurately modeled dynamics derived from MazurkaBL could inform generative models for expressive performance synthesis.
• Fine-Grained Expressive Structure Discovery: Joint modeling of dynamics and metrical hierarchy encourages research into musicological questions regarding phrasing, rubato, and inner variation.

A plausible implication is that the methodological advances validated on MazurkaBL will stimulate broader adoption of multi-task architectures in MIR, and underpin future work in expressivity-driven computational systems.

---

In summary, the MazurkaBL dataset is a cornerstone resource for rhythmic and dynamic modeling in solo piano music, underpinning state-of-the-art multi-task networks that efficiently capture expressive and metrical structure. Its technical properties, evaluation protocols, and significance in both performance analysis and resource-efficient system design establish its centrality in contemporary computational musicology.