In a groundbreaking development, researchers have harnessed the power of large language models (LLMs) to enhance automatic chord recognition, a critical task in music information retrieval (MIR). This innovative approach leverages the reasoning capabilities of LLMs to integrate and refine outputs from various state-of-the-art MIR tools, ultimately improving the accuracy of chord recognition.
The research, led by Chih-Cheng Chang, Bo-Yu Chen, Lu-Rong Chen, and Li Su, introduces a novel framework that positions text-based LLMs as intelligent coordinators. These models process and integrate information from diverse MIR tools, including music source separation, key detection, chord recognition, and beat tracking. By converting audio-derived musical information into textual representations, the LLMs can perform reasoning and correction specifically for chord recognition tasks. This method allows for a more holistic and accurate analysis of musical content, as the LLM can cross-reference and validate information from multiple sources.
The researchers designed a 5-stage chain-of-thought framework that enables GPT-4o, a variant of the GPT model, to systematically analyze, compare, and refine chord recognition results. This framework leverages music-theoretical knowledge to integrate information across different MIR components, ensuring that the final chord recognition output is as accurate as possible. The experimental evaluation on three datasets demonstrated consistent improvements across multiple evaluation metrics, with overall accuracy gains ranging from 1% to 2.77% on the MIREX metric. These results highlight the potential of LLMs to function as integrative bridges in MIR pipelines, facilitating better coordination and information sharing between different tools.
The practical applications of this research are significant for the music and audio production industries. Enhanced automatic chord recognition can improve music transcription services, making it easier for musicians to create sheet music from recordings. It can also aid in music education by providing more accurate tools for analyzing and learning from recorded music. Additionally, this technology can be integrated into music production software to assist composers and producers in creating and refining their work. By improving the accuracy of chord recognition, this research opens up new possibilities for how we interact with and understand musical content. Read the original research paper here.
