Article: Intelligent quantification of piano performance quality via bidirectional long short-term memory networks Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.21 pp.32 - 46 Abstract: Piano performance evaluation is a key component of music education, yet traditional teacher-based assessment is limited by subjectivity, inconsistent standards, delayed feedback, and poor scalability, motivating the need for an objective, automated scoring method. This study proposed a bidirectional long short-term memory network (BiLSTM) with an attention mechanism for automatic piano performance scoring. Five features, including pitch, rhythm, dynamics, pedal usage, and expressiveness, were extracted from MIDI data and reduced from 142 to 100 dimensions before model input. BiLSTM captured bidirectional temporal dependencies, while attention highlighted critical segments for five-level classification. Experiments on 10,000 MIDI samples achieved 0.873 accuracy under five-fold cross-validation, with precision, recall and F1-score of 0.865, 0.858 and 0.861, respectively. Ablation results showed improvements over LSTM (0.081) and BiLSTM (0.031), confirming the effectiveness of attention and strong generalisation on unseen pieces (≥0.851). Inderscience Publishers - linking academia, business and industry through research

Title: Intelligent quantification of piano performance quality via bidirectional long short-term memory networks

Authors: Juanjuan Luo

Addresses: School of Music and Dance, Hunan University of Arts and Science, Changde, 415000, China

Abstract: Piano performance evaluation is a key component of music education, yet traditional teacher-based assessment is limited by subjectivity, inconsistent standards, delayed feedback, and poor scalability, motivating the need for an objective, automated scoring method. This study proposed a bidirectional long short-term memory network (BiLSTM) with an attention mechanism for automatic piano performance scoring. Five features, including pitch, rhythm, dynamics, pedal usage, and expressiveness, were extracted from MIDI data and reduced from 142 to 100 dimensions before model input. BiLSTM captured bidirectional temporal dependencies, while attention highlighted critical segments for five-level classification. Experiments on 10,000 MIDI samples achieved 0.873 accuracy under five-fold cross-validation, with precision, recall and F1-score of 0.865, 0.858 and 0.861, respectively. Ablation results showed improvements over LSTM (0.081) and BiLSTM (0.031), confirming the effectiveness of attention and strong generalisation on unseen pieces (≥0.851).

Keywords: bidirectional long short-term memory network; BiLSTM; piano performance; automatic scoring; artificial intelligence.

DOI: 10.1504/IJRIS.2026.155787

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.21, pp.32 - 46

Received: 28 May 2026
Accepted: 30 Jun 2026
Published online: 13 Aug 2026 *

Title: Intelligent quantification of piano performance quality via bidirectional long short-term memory networks

Keep up-to-date