MQAScore evaluates text-to-music alignment by using a large audio language model to verify five atomic musical dimensions: genre, instrument, mood, vocal, and usage. This method replaces global similarity scores with interpretable, clip-level binary verification.
HOW THIS AFFECTS YOU
●
researcherYou can use this to move beyond global embedding similarity for evaluating generative audio models.
●
designerYou can gain more granular control and feedback when generating music via text prompts.