Integration of Deep Learning in Automatic Music Generation, Aiming at Preserving and Developing Don Ca Tai Tu

Overview

The application of modern technologies is essential for preserving and revitalizing Don ca tai tu, a UNESCO-recognized Intangible Cultural Heritage of Humanity, increasingly overshadowed by the rise of contemporary music genres. However, current efforts primarily reach a limited audience, while younger generations remain largely uninterested in this traditional art form. To address this challenge, we developed a deep learning framework for automatic music generation, allowing users to compose musical pieces infused with the stylistic essence of Don ca tai tu using natural language prompts. We curated a specialized dataset by collecting recordings from multiple online sources, experienced musicians, and cultural experts. The data was meticulously filtered and processed to ensure a clean, high-quality training set. Moreover, we introduce DcttGen, a unified framework to generate high-fidelity, long-form compositions based on text inputs. It achieves this through (1) Music Tokenizer to extract essential musical features from the raw audio waveform efficiently, (2) Autoregressive Transformer to maintain long-term structural coherence in the generated music, and (3) Rectified Flow Transformer to synthesize high-fidelity audio outputs from low sampling rate representations. Furthermore, we redesigned the Chain-of-Thought prompting technique tailored for the Don ca tai tu music, enhancing the resulting compositions' coherence and creativity. With this combination, DcttGen can produce high-quality, well-structured musical pieces compared to other baselines. Experimental results show that DcttGen matches or even surpasses recent methods in various objective metrics. Through this work, we introduce a novel AI-driven approach to the preservation and revitalization of Don ca tai tu.

Competition history

• ISEF 2025 Technology Enhances the Arts · Entry TECA021T

Resources

• ISEF abstract page ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.