Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

YourMT3+: Multi-instrument Music Transcription with Enhanced Transformer Architectures and Cross-dataset Stem Augmentation

Sungkyun Chang, Emmanouil Benetos, Holger Kirchhoff, Simon Dixon

TL;DR

YourMT3+ advances multi-instrument automatic music transcription by integrating a PerceiverTF-based encoder with MT3-style token decoding, enhanced with a mixture-of-experts and a multi-channel decoder to handle incomplete annotations. It introduces online data augmentation, including intra-stem and cross-dataset stem mixing plus pitch-shifting, enabling training across diverse datasets and enabling direct vocal transcription without voice separation front-ends. Across ten public datasets, the approach delivers competitive or superior results to state-of-the-art models, with notable gains from the MoE encoder and multi-channel decoding, though singing transcription and pop-music performance reveal remaining gaps likely due to synthetic-data biases. The work provides fully reproducible code and data pipelines, and highlights practical implications for robust, scalable transcription in real-world music analysis and production contexts.

Abstract

Multi-instrument music transcription aims to convert polyphonic music recordings into musical scores assigned to each instrument. This task is challenging for modeling as it requires simultaneously identifying multiple instruments and transcribing their pitch and precise timing, and the lack of fully annotated data adds to the training difficulties. This paper introduces YourMT3+, a suite of models for enhanced multi-instrument music transcription based on the recent language token decoding approach of MT3. We enhance its encoder by adopting a hierarchical attention transformer in the time-frequency domain and integrating a mixture of experts. To address data limitations, we introduce a new multi-channel decoding method for training with incomplete annotations and propose intra- and cross-stem augmentation for dataset mixing. Our experiments demonstrate direct vocal transcription capabilities, eliminating the need for voice separation pre-processors. Benchmarks across ten public datasets show our models' competitiveness with, or superiority to, existing transcription models. Further testing on pop music recordings highlights the limitations of current models. Fully reproducible code and datasets are available with demos at \url{https://github.com/mimbres/YourMT3}.

YourMT3+: Multi-instrument Music Transcription with Enhanced Transformer Architectures and Cross-dataset Stem Augmentation

TL;DR

YourMT3+ advances multi-instrument automatic music transcription by integrating a PerceiverTF-based encoder with MT3-style token decoding, enhanced with a mixture-of-experts and a multi-channel decoder to handle incomplete annotations. It introduces online data augmentation, including intra-stem and cross-dataset stem mixing plus pitch-shifting, enabling training across diverse datasets and enabling direct vocal transcription without voice separation front-ends. Across ten public datasets, the approach delivers competitive or superior results to state-of-the-art models, with notable gains from the MoE encoder and multi-channel decoding, though singing transcription and pop-music performance reveal remaining gaps likely due to synthetic-data biases. The work provides fully reproducible code and data pipelines, and highlights practical implications for robust, scalable transcription in real-world music analysis and production contexts.

Abstract

Multi-instrument music transcription aims to convert polyphonic music recordings into musical scores assigned to each instrument. This task is challenging for modeling as it requires simultaneously identifying multiple instruments and transcribing their pitch and precise timing, and the lack of fully annotated data adds to the training difficulties. This paper introduces YourMT3+, a suite of models for enhanced multi-instrument music transcription based on the recent language token decoding approach of MT3. We enhance its encoder by adopting a hierarchical attention transformer in the time-frequency domain and integrating a mixture of experts. To address data limitations, we introduce a new multi-channel decoding method for training with incomplete annotations and propose intra- and cross-stem augmentation for dataset mixing. Our experiments demonstrate direct vocal transcription capabilities, eliminating the need for voice separation pre-processors. Benchmarks across ten public datasets show our models' competitiveness with, or superiority to, existing transcription models. Further testing on pop music recordings highlights the limitations of current models. Fully reproducible code and datasets are available with demos at \url{https://github.com/mimbres/YourMT3}.
Paper Structure (15 sections, 1 equation, 2 figures, 3 tables, 1 algorithm)

This paper contains 15 sections, 1 equation, 2 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Relation to Prior Work
  3. Model
  4. Input
  5. Encoder
  6. Output Tokens
  7. Decoder
  8. Data Augmentation
  9. Intra-stem Augmentation
  10. Cross-dataset Stem Augmentation
  11. Pitch-shifting
  12. Experiments
  13. Experimental Setup
  14. Results and Discussion
  15. Conclusion and Future Work

Figures (2)

  • Figure 1: Overview of YourMT3+. (left) Our encoder $E(\cdot)$ takes as input a log mel spectrogram $S$ derived from audio $X$. (center) An auto-regressive decoder $D(\cdot)$ with the language model (LM) head is conditioned by $E(S)$, and output event tokens $Y'$. (right) Cross-dataset stem augmentation, described in Section \ref{['sec:augmentation']}.
  • Figure 2: Instrument Onset F1 on Slakh manilow2019cutting.