create symbolic music representations

This work addresses the limitations of existing symbolic music tokenization approaches, which typically rely on event sequences with irregular time steps and struggle to explicitly model rhythmic regularities. The authors propose a novel tokenization method that uses fixed temporal units—such as beats—as the fundamental building blocks, merging all events sharing the same pitch within a single time step into one token, thereby yielding a sparse piano-roll-like representation. This approach enables explicit alignment of temporal structure for the first time. Integrated with a Transformer architecture, it significantly enhances generation quality, structural coherence, and long-range dependency modeling in tasks such as music continuation and accompaniment generation, while also achieving superior efficiency and rhythmic consistency compared to prevailing event-based tokenization methods.

This paper addresses the problem of automatic MIDI performance quantization into canonical musical notation. Methodologically, it proposes a Transformer model that jointly incorporates beat structure and accent perception: (i) a beat-aligned preprocessing pipeline unifies symbolic representations of performances and scores; (ii) beat-position encoding and accent-aware token embeddings explicitly model hierarchical metrical relationships. Trained on multi-style piano and guitar performance data, the model achieves state-of-the-art performance on the MUSTER benchmark—improving rhythmic quantization accuracy by 4.2 percentage points (absolute) over prior methods, with notably enhanced robustness on complex rhythms and freely interpreted passages. The core contribution lies in deeply integrating beat-level structural priors into the Transformer architecture, enabling end-to-end, interpretable, high-accuracy music quantization.

Conventional MIDI-based absolute-pitch tokenization in symbolic music analysis suffers from limited expressive power and poor interpretability. Method: This paper proposes an interval-driven, range-based note tokenization paradigm that replaces absolute pitch with relative interval relationships as fundamental semantic units, establishing a theoretically consistent and structure-aware music tokenization framework. We design a symbolic tokenization strategy grounded in MIDI interval computation, tailored for Transformer architectures, and evaluate it uniformly across three tasks: key analysis, melody prediction, and harmonic recognition. Results: Experiments demonstrate significant accuracy improvements across all tasks. Moreover, the model’s attention mechanisms explicitly concentrate on musically meaningful structural features—such as stepwise motion and leaps—enabling, for the first time, a systematic integration of music-theoretic insights with neural language modeling capabilities.

In symbolic music generation, fine-grained tokenization ensures high-fidelity modeling at the cost of computational efficiency, whereas compact (compound) tokenization improves decoding speed but hinders intra-token dependency modeling. To address this trade-off, we propose a dynamic positional (DP) token scheduling mechanism that autoregressively unfolds compound tokens during decoding, enabling explicit modeling of internal token structure without additional parameters. DP integrates seamlessly into existing representation frameworks via a delayed scheduling strategy. Experiments on a symphonic MIDI dataset demonstrate that our method significantly enhances musical structural coherence and generation quality over standard compound tokenization, while substantially narrowing the performance gap with fine-grained tokenization. Crucially, DP achieves this improvement without compromising decoding efficiency—thus reconciling high-fidelity modeling with scalable inference.

This work addresses three core tasks—instrument cloning, text-to-instrument synthesis, and text-guided timbre manipulation—without fine-tuning. Methodologically, it introduces a unified, token-based decoder-only Transformer that jointly conditions on MIDI sequences and CLAP-derived text/audio embeddings, generating audio autoregressively at the token level via a neural audio codec (e.g., SoundStream). Crucially, it achieves zero-shot cross-task generalization: a single model supports instrument cloning (from reference audio only), text-driven synthesis (e.g., “violin playing jazz”), and fine-grained timbre editing (e.g., “brighter and softer”). Quantitative and perceptual evaluations demonstrate state-of-the-art performance in audio quality, timbre similarity, and MIDI fidelity. The framework is fully open-sourced, including code, pretrained weights, and interactive demos, validating both its technical efficacy and practical utility.

This work addresses the challenge of simultaneously modeling fine-grained event-level timing and long-range structural coherence in audio-driven music game level generation. To this end, we propose a multimodal sequence-to-sequence approach that conditions on both audio segments and level metadata, employing an event-based tokenized representation that explicitly encodes beat-aligned actions and their relative temporal offsets—thereby overcoming the limitations of conventional frame-level representations. Built upon a Transformer architecture, our model jointly learns the alignment between audio and event sequences to generate levels with precise rhythmic fidelity. Experimental results demonstrate that the proposed method significantly outperforms frame-level baselines on event-level evaluation metrics and enables systematic analysis of how audio cues contribute to rhythm-aligned prediction.

This work addresses the limited controllability of existing symbolic music generation methods, which typically rely on from-scratch synthesis and struggle to support explicit local editing. The paper introduces the first explicit editing framework for symbolic music, reframing generation as a draft-editing process. Built upon a BEAT-based rhythmic grid anchoring scheme, the framework unifies three editing mechanisms—token-wise sequence labeling, iterative accompaniment refinement, and post-hoc token infilling—within a single pre-trained backbone model to enable efficient inference. Experimental results demonstrate that the proposed approach outperforms both autoregressive and diffusion models across three distinct editing tasks, achieving inference latency under 100 milliseconds while significantly improving both generation accuracy and perceptual audio quality. These findings highlight a strong correlation between music representation design and editing efficacy.

This study addresses the lack of systematic evaluation of music tokenization’s impact in text-to-symbolic-music generation. By fixing the pretrained model, training data, and decoding strategy while varying only the symbolic representation, the authors systematically assess seven tokenization schemes in terms of distributional fidelity. Their findings reveal that the choice of musical representation—not model scale—is the primary determinant of generation quality. They propose PMT (Performance-level Music Tokenization), a high-fidelity representation that enables even a 0.8B-parameter model to achieve an FMD score of 159, substantially outperforming 27B-parameter models using beat-grid representations (FMD: 272–286). Additionally, lightweight decoding constraints double instrument F1 and tonality accuracy while preserving distributional consistency.