ZGUERGUER, Amen Allah (2024) Building a large music dataset from existing datasets PRE - Research Project, ENSTA.
Full text not available from this repository.
Abstract
Automatic music captioning is a task within the field of Music Information Retrieval that involves generating human-like descriptive text for a piece of music. Training Music captioning models requires a large amount of high-quality annotated data, which is often scarce. The goal behind my internship is to address this data scarcity for this task by building a large music dataset that contains music-text pairs. This dataset is constructed from existing tagging-music datasets, leveraging a tag-to-description approach using a Large Language Model, Mistral. We carried out an objective evaluation of the generated captions using appropriate metrics and our method achieved better results than other methods.
Item Type: | Thesis (PRE - Research Project) |
---|---|
Uncontrolled Keywords: | Music captioning, Music Information Retrieval, Music Dataset, Music auto- tagging, Large Language Model, Metrics |
Subjects: | Information and Communication Sciences and Technologies |
ID Code: | 10081 |
Deposited By: | Amen allah ZGUERGUER |
Deposited On: | 02 sept. 2024 18:03 |
Dernière modification: | 02 sept. 2024 18:03 |
Repository Staff Only: item control page