Pseudo-Supervised Training Improves Unsupervised Melody Segmentation

Stefan Lattner, Carlos Eduardo Cancino-Chacón, Maarten GrachtenIJCAI 2015

Introduction

Melody segmentation—the process of dividing a continuous stream of musical notes into meaningful units like phrases or motifs—is a fundamental task in music cognition and computational musicology. While humans perform this "chunking" effortlessly, automating it remains a challenge. A prominent theory suggests that boundaries are perceived at points of high Information Content (IC), where a note is particularly unexpected given its context. However, raw statistical estimates of IC are often noisy. This paper presents a "pseudo-supervised" approach to refine these estimates, significantly improving segmentation accuracy without requiring human-labeled training data.

The Pseudo-Supervised Pipeline

The core innovation is a three-stage process that uses a generative model to "teach" a predictive model.

  1. Generative Modeling: A Restricted Boltzmann Machine (RBM) is trained on a large corpus of melodies (the Essen Folk Song Collection). It learns the statistical structure of note sequences (n-grams) based on features like pitch intervals and timing.
  2. Generating Pseudo-Targets: The trained RBM calculates the IC for every note. Since exact calculation is computationally expensive, Monte Carlo techniques (Gibbs sampling) are used to estimate these values, creating a "Boundary Strength Profile" (BSP).
  3. Pseudo-Supervised Training: Instead of using the RBM's noisy IC estimates directly, the researchers use them as "pseudo-labels" to train a standard Feed-Forward Neural Network (FFNN). The FFNN learns to map raw musical features to these IC values.

This "bootstrapping" step is crucial. Because the FFNN is forced to generalize across the dataset, it effectively filters out the local noise and outliers present in the RBM's raw estimates—a process the authors describe as context-sensitive smoothing.

Results and Analysis

The models were evaluated on the German folk song subset of the Essen Folk Song Collection, using human-annotated phrase markers as the ground truth. Performance was measured using the F1 score, which balances precision and recall.

F1 scores across different N-gram lengths, showing the performance boost from pseudo-supervision (PS).

As shown in the figure above, the pseudo-supervised model (RBM+DO+PS) consistently outperforms the baseline RBM and the RBM with dropout (DO) across all tested N-gram lengths. The performance gain is attributed to the FFNN's ability to act as an "entropy regularizer," ignoring the statistical noise that often plagues raw generative models.

Performance comparison of the proposed RBM10+DO+PS model against established baselines.

The final results place the proposed unsupervised method on par with established rule-based models like LBDM (F1 = 0.63). While it does not yet surpass Grouper (F1 = 0.66), which incorporates sophisticated domain-specific knowledge like musical parallelism, it represents a significant advancement for purely statistical, "agnostic" models.

Conclusion

The success of pseudo-supervised training in melody segmentation highlights the power of using neural networks to "smooth" noisy statistical signals. By learning to predict its own generative estimates, the system achieves a more robust understanding of musical structure. Future work may explore more complex architectures like Recurrent Neural Networks (RNNs) to better capture long-term temporal dependencies or apply this pseudo-supervised strategy to other domains where ground-truth labels are difficult to obtain.

Summary updated

BibTeX
@inproceedings{lattner2015pseudo,
  author = {Stefan Lattner and Carlos Eduardo Cancino-Chacón and Maarten Grachten},
  title = {{Pseudo-Supervised Training Improves Unsupervised Melody Segmentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}

Loading related content...

Figure