3. Music Information Extraction

3.1. Lyrics

Any lyric-related tasks can be effectively solved by LLMs, as they are also a language task. For example, in my opinion, lyric-based mood recognition can outperform any audio-based methods in the past, and perhaps even in the future.

Perhaps we should go further - for example, singer diarization (to classify phrases into a singer class when there are multiple singers in a song) can be done with help of LLMs.

3.2. Metadata

Although LLMs are not necessarily factual, their embeddings and understanding of some popular pronouns are pretty good. This means we may rely on LLMs for artist or album understanding when they’re popular enough.

3.3. Other Unstructured Data

Text data such as album reviews and artist biographies can be used as a target material to utilize LLMs and perform music information retrieval.