Multimodal Pretraining for Music Audio

Summary

Data can be expressed in various forms, each potentially encoded through diverse means. For instance, we might encounter audio data paired with descriptive texts about their lyrics. Modern systems leverage, if available, the different sources of information and outperform, under certain conditions, their single-modal counterparts. In such multimodal settings, each modality encapsulates a distinct aspect of the underlying semantics of the data and has a supplementary role. Data can also be limited and without annotations related to the task at hand. In such cases, transfer learning and pretraining could be two techniques that enhance the performance of the models. In this thesis, we explore various unsupervised pretraining techniques while evaluating them on a supervised downstream task. Our goal is to train a model that can extract meaningful features and be further finetuned to any new task. We use LLMs to create pseudo-captions that describe the sentiment and the theme of the lyrics, from a large pool of non-annotated audio. We then perform a pretraining step, where we learn a multimodal coordinated space between the audio signals and these pseudo-captions. Then, we finetune our model on an annotated dataset, where only the audio modality is available. We highlight the ability of such models to deliver adequate performance in few-shot learning settings, the incorporation of LLMs into the pretraining step, and the importance of learning a shared semantic space for information originating from different modalities.