LLMs in Music Composition
Data Skeptic
Mon, August 28, 2023
Podchat Summary

Exploring the Future of Music Composition with Machine Learning

In this episode, Carlos Hernández-Olivan, a PhD student at the University of Zaragoza, delves into the fascinating world of music composition and its intersection with machine learning. Carlos discusses the potential of new technologies, such as digital audio workstations, in revolutionizing the way composers create music. While some composers may still prefer traditional methods like staff paper or a piano, digital audio workstations offer the advantage of composing and listening simultaneously.

One of the most exciting applications of machine learning in music composition is its ability to assist composers by providing new ideas and speeding up the composition process. Carlos explains how machine learning algorithms can generate music that is stylistically accurate and can overcome creative blocks. Unlike random music generators, machine learning models, particularly those based on the transformer architecture, can process MIDI files and convert them into text tokens that can be input to the model.

However, evaluating machine-generated music poses unique challenges. Traditional metrics struggle to measure music quality and emotion accurately. Carlos explores the use of objective and subjective evaluations but acknowledges the need for improvement in both areas. Deep learning models have the potential to go beyond simple introductory compositions and introduce new ideas and flourishes that can be regarded as creative.

The availability of large datasets and the size of the models used in music generation play a crucial role in determining the quality and coherence of the generated music. Carlos emphasizes the importance of collaboration between composers and researchers in developing music generation models that focus on important musical elements, such as motifs, and produce high-quality compositions.

Carlos also discusses the current state of the art in music generation, highlighting how machine-generated music can often fool listeners into thinking it was composed by humans, especially in symbolic music generation with good synthesis and post-processing. However, several open questions remain, including the creativity of machine-generated music, the optimal model size for generating coherent music, the availability of diverse datasets for different music genres, and the potential impact of AI technologies on the profession of musicians.

Looking ahead, Carlos shares his plans to continue working on music generation, particularly in the symbolic domain, and exploring the intricate relationship between music and emotions. Listeners can follow Carlos on Twitter @CarlosAndOliv to stay updated on his research and insights.

Original Show Notes

In this episode, we are joined by Carlos Hernández Oliván, a Ph.D. student at the University of Zaragoza. Carlos’s interest focuses on building new models for symbolic music generation.

Carlos shared his thoughts on whether these models are genuinely creative. He revealed situations where AI-generated music can pass the Turing test. He also shared some essential considerations when constructing models for music composition.

Made with ☕️ in SF/SD.
© 2023 Spyglass Search, Inc.