A Feature-Enriched BiLSTM Architecture for Automated Music Composition

• Published in: Indian journal of science and technology Vol. 18; no. 31; pp. 2527 - 2538
• Main Authors: Kapoor, Bhumik, Kaur, Gurleen, Chawla, Khushi, Saini, Utkarsh, Gupta, Monica, Kaur, Surinder
• Format: Journal Article
• Language: English
• Published: 23.08.2025
• ISSN: 0974-6846; 0974-5645
• DOI: 10.17485/IJST/v18i31.879

Objectives: To create suitable framework for BiLSTM neural networks to learn various musical styles after analyzing musical files and to generate natural melodies. Methods: This study has developed a novel program. Firstly, it extracts important features from the NSynth data set, using Mel spectrograms and Mel-frequency cepstral coefficients to review all the given audio material. Each feature is given a unique number which is validated, divided evenly among groups and shuffled to be unbiased. The specific sequential arrangement of audio features makes it easier for BiLSTM network to detect the time sequence in the data. Training is done by using batch normalization and a 0.6 dropout rate, adjustable learning rates, early stopping and the Huber loss function. Findings: The model is versatile enough to create MIDI files with acceptable tunes for different groups of instruments, regardless of the music style. Music can mimic natural situations by mixing quick tempo changes with choices of limited sounds that are properly in control. Novelty: It combines BiLSTM network with different audio feature extraction techniques from the NSynth dataset to suggest a different approach for generating music. Merging melodic music with machine learning results into a structure that offers a musician, flexibility and structure. Phase- based generation helps create sections in music while retaining the special sound of each musical group. It supports musicians by offering assistance for creativity, fit for the needs of distinct instruments and music genres. Keywords: BiLSTM, Music Generation, Deep Learning, NSynth, MIDI Synthesis, Temporal Modelling, AI Composition