Language-Based Audio Retrieval with Converging Tied Layers and Contrastive Loss
This work addresses audio retrieval from text queries, an incremental improvement for multimedia and audio processing applications.
The paper tackles language-based audio retrieval by introducing a tied-layer architecture with contrastive loss, which significantly outperforms the baseline model while requiring low training memory and no fine-tuning of pretrained models.
In this paper, we tackle the new Language-Based Audio Retrieval task proposed in DCASE 2022. Firstly, we introduce a simple, scalable architecture which ties both the audio and text encoder together. Secondly, we show that using this architecture along with contrastive loss allows the model to significantly beat the performance of the baseline model. Finally, in addition to having an extremely low training memory requirement, we are able to use pretrained models as it is without needing to finetune them. We test our methods and show that using a combination of our methods beats the baseline scores significantly.