Article: Music score recognition and stylised music generation based on multi-scale feature fusion CRNN and GAN-transformer Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.90 pp.1 - 26 Abstract: In order to solve the frequent loss of small symbols in complex music scores and the serious interference of five line staff backgrounds, a research proposes an end-to-end architecture that integrates music score recognition and style generation. The recognition stage adopts a multi-scale feature fusion convolutional recurrent neural network, combined with Haar wavelet transform to preserve edge details, and utilises note focusing module and quasi recurrent neural network for efficient transcription. The generation stage adopts a hierarchical cascade strategy, combining generative adversarial networks and Transformers to capture structural and spectral details, and then synthesising waveforms through WaveNet voice encoders. The results showed that the symbol accuracy on the DeepScores V2 dataset was 96.82%, and the symbol error rate under noise interference was only 5.15%. The research results provide effective technical support for the audio-visual conversion from visual music scores to auditory audio. Inderscience Publishers - linking academia, business and industry through research

Title: Music score recognition and stylised music generation based on multi-scale feature fusion CRNN and GAN-transformer

Authors: Fan Yang; Yijing Chen

Addresses: Nanhai Conservatory of Music, Haikou University of Economics, Haikou 571127, China ' Nanhai Conservatory of Music, Haikou University of Economics, Haikou 571127, China

Abstract: In order to solve the frequent loss of small symbols in complex music scores and the serious interference of five line staff backgrounds, a research proposes an end-to-end architecture that integrates music score recognition and style generation. The recognition stage adopts a multi-scale feature fusion convolutional recurrent neural network, combined with Haar wavelet transform to preserve edge details, and utilises note focusing module and quasi recurrent neural network for efficient transcription. The generation stage adopts a hierarchical cascade strategy, combining generative adversarial networks and Transformers to capture structural and spectral details, and then synthesising waveforms through WaveNet voice encoders. The results showed that the symbol accuracy on the DeepScores V2 dataset was 96.82%, and the symbol error rate under noise interference was only 5.15%. The research results provide effective technical support for the audio-visual conversion from visual music scores to auditory audio.

Keywords: convolutional recurrent neural network; CRNN; music score recognition; generative adversarial network; GAN; transformer; music generation.

DOI: 10.1504/IJICT.2026.155733

International Journal of Information and Communication Technology, 2026 Vol.27 No.90, pp.1 - 26

Received: 05 Feb 2026
Accepted: 06 May 2026
Published online: 11 Aug 2026 *

Title: Music score recognition and stylised music generation based on multi-scale feature fusion CRNN and GAN-transformer

Keep up-to-date