The timbral analysis from spectrographic features of popular music sub-genres (or micro-genres) presents unique challenges to the field of the computational auditory scene analysis, which is caused by the adjacencies among sub-genres and the complex sonic scenes from sophisticated musical textures and production processes. This paper presents a timbral modeling tool based on a modified deep learning natural language processing model. It treats the time frames in spectrograms as words in natural languages to explore the temporal dependencies. The modeling performance metrics obtained from the fine-tuned classifier of the modified Deep Bidirectional Encoder Representations from Transformers (BERT) model show strong semantic modeling performances with different temporal settings. Designed as an automatic feature engineering tool, the proposed framework provides a unique solution to the semantic modeling and representation tasks for objectively understanding of subtle musical timbral patterns from highly similar musical genres.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=21063
(1597KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
