YourMT3+: Multi-instrument Music Transcription with Enhanced Transformer Architectures and Cross-dataset Stem Augmentation (2407.04822v3)

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}.

• The paper introduces YourMT3+, a model utilizing hierarchical attention and a Mixture of Experts to boost multi-instrument transcription performance.
• It features a multi-channel decoder that effectively addresses incomplete annotations by assigning distinct tasks to different channels.
• The cross-dataset stem augmentation strategy enriches training data diversity, mitigating data scarcity and enhancing transcription robustness.

Multi-instrument Music Transcription Using Enhanced Transformer Architectures

The paper "YourMT3+: Multi-instrument Music Transcription with Enhanced Transformer Architectures and Cross-dataset Stem Augmentation" examines advanced methodologies to tackle the complex task of multi-instrument music transcription. This research introduces YourMT3+, a suite of models built upon the MT3 framework and enhanced with novel methods to improve transcription accuracy, even in the presence of data scarcity.

Core Contributions

1. Enhanced Transformer Architectures:
   • The authors incorporate a hierarchical attention mechanism within the encoder, specifically using PerceiverTF's spectral cross-attention strategy. This enhancement improves the model's capacity to process and differentiate the complex and overlapping sound sources inherent in multi-instrument music.
   • A novel integration of a Mixture of Experts (MoE) within the encoder is implemented. MoE facilitates task-specific processing, thereby improving model flexibility and performance across varied input data types.
2. Multi-channel Decoder:
   • The paper proposes a multi-channel decoding architecture, which significantly benefits the training process, allowing the model to seamlessly handle datasets with incomplete annotations. This decoder assigns different instrument group tasks across channels, enhancing the transcription accuracy via better data handling.
3. Cross-dataset Stem Augmentation:
   • The study introduces a cross-dataset stem augmentation strategy, leveraging both intra- and cross-stem mixing to address the issues of data inadequacy. This technique augments datasets by selectively combining musical stems from different sources, creating a more diverse and comprehensive training regime.

Evaluation and Outcomes

The models were benchmarked across multiple datasets, showing superior results compared to existing methods. Notably, the YourMT3+ suite outperformed the original MT3 system and demonstrated remarkable robustness in multi-instrument transcription tasks. However, certain challenges were highlighted, such as performance variability on commercial pop music datasets and minor timing discrepancies in voice transcription, illustrating areas for future exploration.

Implications and Future Directions

From a practical standpoint, the findings and methodologies outlined in this paper hold significant promise for commercial applications where accurate music transcription is crucial, such as in music production and automated accompaniment systems. Furthermore, the success of both MoE and the hierarchical attention model paves the way for potential adaptations in other domains requiring granular auditory recognition.

For future developments, exploring more comprehensive datasets, particularly those covering commercial music diversity, could alleviate current transcription biases. Additionally, optimizing pitch-shifting parameters within the training regime could address octave error challenges identified in the experimental results.

In conclusion, the proposed YourMT3+ model suite marks a pertinent advancement in the field of music information retrieval, pushing the boundaries of current multi-instrument transcription methods and setting a benchmark for subsequent research initiatives.