Uploaded September 2026 | Updated November 2026, 17 hours ago
How do you evaluate a generative AI music system like Suno, without asking thousands of humans to listen to its output?
In this deep dive, we explore Fréchet Audio Distance (FAD), one of the most widely used objective metrics for evaluating generative audio and music models. We build the intuition from embedding distributions and Gaussian approximations, unpack the mathematics behind FAD, and implement it from scratch with Python using synthetic examples.
KEY TOPICS
- Evaluating generative AI music and audio models
- Fréchet Audio Distance (FAD) and Fréchet Inception Distance (FID)
- Audio embeddings and representation models
- Gaussian distributions and multivariate Gaussians
- Mean, variance, covariance, and covariance matrices
- FAD formula: mean mismatch + covariance mismatch
- Trace, (a pinch of eigenvalues, eigenvectors, and principal directions)
- Mode collapse and insufficient diversity
- Python implementation of FAD from scratch
- Limitations of FAD: representation dependence, creativity, and novelty
CONSULTING:
🚀 AI Music + Audio Consulting: valeriovelardoadvisor.com
📩 Get my AI Music content in your inbox for free: valeriovelardo.substack.com
JOIN THE SOUND OF AI COMMUNITY ON SLACK:
valeriovelardo.com/the-sound-of-ai-community
CODE + SLIDES
Get code and slides (math + presentation) from the FAD GitHub repo: github.com/musikalkemist/fad
CHAPTERS:
0:00 Intuition
21:22 Deep math dive
1:48:00 Code implementation
How do you evaluate a generative AI music system like Suno, without asking thousands of humans to listen to its output?
In this deep dive, we explore Fréchet Audio Distance (FAD), one of the most widely used objective metrics for evaluating generative audio and music models. We build the intuition from embedding distributions and Gaussian approximations, unpack the mathematics behind FAD, and implement it from scratch with Python using synthetic examples.
KEY TOPICS
- Evaluating generative AI music and audio models
- Fréchet Audio Distance (FAD) and Fréchet Inception Distance (FID)
- Audio embeddings and representation models
- Gaussian distributions and multivariate Gaussians
- Mean, variance, covariance, and covariance matrices
- FAD formula: mean mismatch + covariance mismatch
- Trace, (a pinch of eigenvalues, eigenvectors, and principal directions)
- Mode collapse and insufficient diversity
- Python implementation of FAD from scratch
- Limitations of FAD: representation dependence, creativity, and novelty
CONSULTING:
🚀 AI Music + Audio Consulting: valeriovelardoadvisor.com
📩 Get my AI Music content in your inbox for free: valeriovelardo.substack.com
JOIN THE SOUND OF AI COMMUNITY ON SLACK:
valeriovelardo.com/the-sound-of-ai-community
CODE + SLIDES
Get code and slides (math + presentation) from the FAD GitHub repo: github.com/musikalkemist/fad
CHAPTERS:
0:00 Intuition
21:22 Deep math dive
1:48:00 Code implementation










