Pianofi: Bridging Transcription and Arrangement in Audio-to-Piano Cover Generation

Abstract

Audio-to-piano cover generation is a powerful technique bridging note based transcription with harmonic understanding of songs. This paper introduces Pianofi; a unified framework for cover generation spanning three SOTA paradigms: AMT-APC, PiCoGen, and Etude. We first construct a training and evaluation pipeline and compare models to clarify trade-offs between baselines. We then introduce enhancements including LoRA adapters, reranking, improved lead-sheet extraction, and targeted data augmentation, yielding upgraded variants. Finally, we explore cross-paradigm ensembling to combine transcription accuracy and harmonic structure regularization. Our results demonstrate consistent F1 and Qmax improvements and more stable piano arrangements, highlighting the benefits of bridging transcription and symbolic generation. The app is deployed at pianofi.ca