Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Analyzing Byte-Pair Encoding on Monophonic and Polyphonic Symbolic Music: A Focus on Musical Phrase Segmentation

Dinh-Viet-Toan Le, Louis Bigo, Mikaela Keller

TL;DR

The findings show that the BPE training process is highly dependent on the instrumentation and that BPE “supertokens” succeed in capturing abstract musical content, and in a musical phrase segmentation task, BPE notably improves performance in a polyphonic setting, but enhances performance in monophonic tunes only within a specific range of BPE merges.

Abstract

Byte-Pair Encoding (BPE) is an algorithm commonly used in Natural Language Processing to build a vocabulary of subwords, which has been recently applied to symbolic music. Given that symbolic music can differ significantly from text, particularly with polyphony, we investigate how BPE behaves with different types of musical content. This study provides a qualitative analysis of BPE's behavior across various instrumentations and evaluates its impact on a musical phrase segmentation task for both monophonic and polyphonic music. Our findings show that the BPE training process is highly dependent on the instrumentation and that BPE "supertokens" succeed in capturing abstract musical content. In a musical phrase segmentation task, BPE notably improves performance in a polyphonic setting, but enhances performance in monophonic tunes only within a specific range of BPE merges.

Analyzing Byte-Pair Encoding on Monophonic and Polyphonic Symbolic Music: A Focus on Musical Phrase Segmentation

TL;DR

The findings show that the BPE training process is highly dependent on the instrumentation and that BPE “supertokens” succeed in capturing abstract musical content, and in a musical phrase segmentation task, BPE notably improves performance in a polyphonic setting, but enhances performance in monophonic tunes only within a specific range of BPE merges.

Abstract

Byte-Pair Encoding (BPE) is an algorithm commonly used in Natural Language Processing to build a vocabulary of subwords, which has been recently applied to symbolic music. Given that symbolic music can differ significantly from text, particularly with polyphony, we investigate how BPE behaves with different types of musical content. This study provides a qualitative analysis of BPE's behavior across various instrumentations and evaluates its impact on a musical phrase segmentation task for both monophonic and polyphonic music. Our findings show that the BPE training process is highly dependent on the instrumentation and that BPE "supertokens" succeed in capturing abstract musical content. In a musical phrase segmentation task, BPE notably improves performance in a polyphonic setting, but enhances performance in monophonic tunes only within a specific range of BPE merges.
Paper Structure (9 sections, 4 figures)

This paper contains 9 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Subword tokenization in symbolic music
  3. Analyzing music BPE
  4. Comparing text and music BPEs
  5. Musical content carried by supertokens
  6. Evaluating BPE on musical phrase segmentation
  7. Musical phrase segmentation
  8. Experiments
  9. Conclusion

Figures (4)

  • Figure 1: (Top) Frequency of the created supertokens through the vocab size increasing with the BPE steps, for different styles of music and multilingual text data. (Bottom) Average length of already created supertokens through BPE iterations for musical and text data. The initial vocabulary size of each tokenization is indicated.
  • Figure 2: (Top) First most common start-of-phrase supertoken from Mozart's K.25 and Beethoven's WoO.68. (Bottom) 9-long common ending supertoken (10th most common) from Beethoven's WoO.73 and Mozart's K.179. The tokenization is Structured + intervals.
  • Figure 3: f1-score for start-of-phrase classification on the polyphonic (top) and monophonic dataset (bottom).
  • Figure 4: Ratio of supertokens containing $n$<Pitch> atomic elements in the vocabulary for each number of BPE merges.