Xi'an Jiaotong University, China
Singapore University of Technology and Design, Singapore

Conditional music generation offers significant advantages in terms of user convenience and control, presenting notable potential in AI-generated content research. When building conditional generative systems for multi-instrument popular songs, two main challenges typically arise: the insufficient fidelity of conditions and the poor quality of the generated music. To address these issues, we first propose multi-view features across time and instruments as high-fidelity and fine-grained conditions. In addition, an efficient music representation called REMI_Track(BPE) is designed to convert multitrack music into multiple music sequences. Subsequently, we release BandCondiNet, a conditional model based on parallel Transformers, designed to process the multiple music sequences and generate high-quality multitrack samples conditioned on the provided multi-view features. Specifically, BandCondiNet incorporates two specialized modules: Structure Enhanced Attention (SEA) to strengthen the resulting musical structure, and Cross-Track Transformer (CTT) to enhance inter-track harmony. Experimental results tested on two popular music datasets of different lengths demonstrate that the proposed BandCondiNet outperforms other conditional music generation models on most objective metrics in terms of fidelity and inference speed, and shows great robustness in generating long music samples. The subjective evaluations further show BandCondiNet trained on short datasets performs comparable to state-of-art models, while outperforming them significantly training on longer datasets.

We conduct experiments on a popular music subset of the LakhMIDI dataset(LMD), which is the largest publicly available symbolic music dataset that contains multiple instruments.

The source code is available at here (To be updated).

Note: In the following MIDI players, the Melody track, the Drum sets track are respectively visualized in magenta and grey.

For reaching me by email, please use: CHINGLOHSIU[AT]gmail[DOT]com or visit my Homepage.
The MIDI Web Player is provided by html-midi-player
Last updated:
Page views: