Neo-Riemannian Cycle Detection with Weighted Finite-State Transducers

Jonathan Bragg, Elaine Chew, Stuart M. Shieber

Introduction

In the study of late Romantic harmony, traditional functional analysis often fails to capture the logic of composers like Franz Schubert or Richard Wagner. These composers frequently employed harmonic progressions that prioritize the voice-leading relationships between individual triads over their relationship to a central tonic key. Neo-Riemannian theory (NRT) provides a framework for analyzing these progressions through transformations such as Parallel (P), Leading-tone (L), and Relative (R). A key interest in NRT is the identification of "cycles"—sequences of chords generated by the iterative application of these transformations.

Methodology: The Noisy Channel Model

Detecting these cycles in actual musical scores is complicated by "musical noise." Composers rarely adhere to strict mathematical cycles; they often insert passing chords or omit steps in a sequence. This research addresses this challenge by framing music analysis as a "noisy channel" problem, similar to techniques used in natural language processing. The observed score is treated as a corrupted version of an underlying, theoretically perfect cycle.

To implement this model, the researchers utilized Weighted Finite-State Transducers (WFSTs). The system is defined by the composition of three primary FSTs:

  1. Score FST ($S$): Encodes the sequence of chords extracted from the musical score.
  2. ScoreEdit FST ($E$): Models the potential "noise" by assigning costs to edit operations such as insertions and deletions.
  3. Cycles FST ($C$): Defines the formal grammar of valid neo-Riemannian cycles.
The ScoreEdit FST modeling deletions (D) and insertions (I) to account for musical noise.

The ScoreEdit FST (Figure 6) allows the system to navigate discrepancies between the theoretical model and the actual score. By assigning weights to these edits, the Viterbi algorithm can identify the path with the lowest cumulative cost, representing the most probable cycle present in the music.

The Cycles FST encoding the grammar of specific neo-Riemannian LP cycles.

The Cycles FST (Figure 4) specifically encodes the sequences of transformations that constitute a cycle, such as the alternating $L$ and $P$ transformations. This structural constraint ensures that the detected patterns align with NRT principles.

Experimental Setup and Results

The model was tested against a manually annotated dataset of four major works by Schubert. The preprocessing stage involved reducing complex harmonies to their constituent major or minor triads, ignoring extended chords to fit the NRT framework.

The system achieved a recall rate of 1.0, successfully identifying every cycle previously noted by human analysts. However, the precision was 0.18, indicating that the model identified many mathematically valid cycles that the human analyst did not consider musically significant.

Discussion and Conclusion

The discrepancy between the model's high recall and low precision highlights a fundamental tension in music theory. While the WFST approach can exhaustively find every theoretical cycle, human analysts likely rely on additional parameters—such as rhythm, meter, and phrasing—to determine which cycles are "salient." This suggests that the model reveals a critique of the theory itself: mathematically valid cycles are ubiquitous, and human judgment is required to filter for musical relevance.

Future work could involve incorporating multi-modal data to better approximate human musical intuition and testing the model on a larger corpus of Romantic music, such as the works of Wagner or Bruckner. Nevertheless, the success of the WFST framework demonstrates a robust method for formalizing and automating complex harmonic analysis in "noisy" musical environments.

Summary updated

BibTeX
@inproceedings{bragg2011neo,
  author = {Jonathan Bragg and Elaine Chew and Stuart M. Shieber},
  title = {{Neo-Riemannian Cycle Detection with Weighted Finite-State Transducers}},
  booktitle = {Ismir},
  year = {2011},
}
Figure