MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training

2023-05-31

International Conference on Learning Representations

AI Summary

To address weak pitch and tonality modeling in music audio understanding and the lack of music-aware priors in existing self-supervised learning (SSL) methods, this paper introduces MERT, a music-specific large-scale self-supervised pretraining framework. Its core innovation is a novel dual-teacher mechanism—combining acoustic modeling via RVQ-VAE and music-perceptual modeling via Constant-Q Transform (CQT)—to enable stable, high-fidelity masked language modeling (MLM) pseudo-label generation. This design overcomes the training instability prevalent in audio foundation models, supporting scalable pretraining across 95M–330M parameter regimes. Evaluated on 14 diverse music understanding tasks, MERT achieves state-of-the-art performance across the board, significantly outperforming general-purpose speech/audio SSL approaches. It establishes the first self-supervised paradigm for music representation learning that jointly ensures acoustic fidelity and music-theoretic consistency.

Abstract

Self-supervised learning (SSL) has recently emerged as a promising paradigm for training generalisable models on large-scale data in the fields of vision, text, and speech. Although SSL has been proven effective in speech and audio, its application to music audio has yet to be thoroughly explored. This is partially due to the distinctive challenges associated with modelling musical knowledge, particularly tonal and pitched characteristics of music. To address this research gap, we propose an acoustic Music undERstanding model with large-scale self-supervised Training (MERT), which incorporates teacher models to provide pseudo labels in the masked language modelling (MLM) style acoustic pre-training. In our exploration, we identified an effective combination of teacher models, which outperforms conventional speech and audio approaches in terms of performance. This combination includes an acoustic teacher based on Residual Vector Quantisation - Variational AutoEncoder (RVQ-VAE) and a musical teacher based on the Constant-Q Transform (CQT). Furthermore, we explore a wide range of settings to overcome the instability in acoustic language model pre-training, which allows our designed paradigm to scale from 95M to 330M parameters. Experimental results indicate that our model can generalise and perform well on 14 music understanding tasks and attain state-of-the-art (SOTA) overall scores.

Problem

• self-supervised learning
• music audio understanding
• pitch and tone recognition

Innovation

• Self-supervised Learning
• Music Understanding
• Pre-training Scaling