Enhance audio generation controllability through representation similarity regularization
This work addresses controllability issues in audio generation for applications like music and sound synthesis, but it is incremental as it builds on existing classifier-free guidance methods.
The paper tackles the problem of poor alignment between audio and text representations in language model-based audio generation by introducing representation similarity regularization during training, which improves objective metrics and human perception for audio generation tasks.
This paper presents an innovative approach to enhance control over audio generation by emphasizing the alignment between audio and text representations during model training. In the context of language model-based audio generation, the model leverages input from both textual and audio token representations to predict subsequent audio tokens. However, the current configuration lacks explicit regularization to ensure the alignment between the chosen text representation and the language model's predictions. Our proposal involves the incorporation of audio and text representation regularization, particularly during the classifier-free guidance (CFG) phase, where the text condition is excluded from cross attention during language model training. The aim of this proposed representation regularization is to minimize discrepancies in audio and text similarity compared to other samples within the same training batch. Experimental results on both music and audio generation tasks demonstrate that our proposed methods lead to improvements in objective metrics for both audio and music generation, as well as an enhancement in the human perception for audio generation.