InvFlowFD: A New Reference-Free and Background-Set-Free Music Quality Metric
A new metric named InvFlowFD has been developed by researchers to assess the perceptual quality of music without requiring reference audio or background sets. Unlike conventional reference-free techniques, which depend on a set of clean audio samples for statistical aggregation, InvFlowFD utilizes only a pre-trained Flow Matching backbone for quality estimation. This method employs unconditional Flow Matching inversion through basic Euler integration, enabling it to identify various artificial distortions and effectively rank music generation models based on human perceptual assessments. By comparing a set of inverted samples to the prior distribution, InvFlowFD's effectiveness was confirmed through quantitative evaluations and comprehensive human studies. Findings show a strong correlation between InvFlowFD and human perceptions of sound distortion. The paper can be found on arXiv with the identifier 2608.04142 and is categorized as 'cross'.
Key facts
- InvFlowFD is a new metric for music perceptual quality.
- It is reference-free and background-set-free.
- It uses a pre-trained Flow Matching backbone.
- It performs unconditional Flow Matching inversion via Euler integration.
- It detects artificial distortions and ranks music generation models.
- It compares inverted samples to the prior distribution.
- Evaluated quantitatively and with a human study.
- Highly correlated with human perception of sound distortion.
Entities
Institutions
- arXiv