ModelCards
← All benchmarks

Text relevance (REL, listener MOS)

We haven't written a description for this benchmark yet; see the card notes below for how each maker ran it.

Your workloadRAG over docs · 12K in · 600 out · 50K requests/mo · 30% cached
  1. 1Stable Audio 2.5 · Stability AI
    4.3 MOS
  2. 2Stable Audio 3.0 Medium · Stability AI
    4.25 MOS
  3. 3Stable Audio 3.0 · Stability AI
    3.8 MOS

“Best value”: no other model here reports a higher score at a lower cost for your workload. Scores are self-reported by each maker with different settings, so treat close calls as ties.

How each maker reported it

  • Stable Audio 2.5 Stable Audio 3 paper (arXiv 2605.17991), Tables 3–4

    The 2.5 launch post has no benchmark numbers. These are Stability's own later comparisons in the Stable Audio 3 paper. FAD lower is better; CLAP higher is better; OVL, REL and MUS are mean opinion scores from 14 listeners; time is seconds on an H200.

  • Stable Audio 3.0 Medium paper, Tables 3–5

    Stability's own results. FAD (lower is better) and CLAP (higher is better) use LAION-CLAP embeddings. OVL, REL and MUS are mean opinion scores from 14 listeners. Time is wall-clock seconds on an H200. Stable Audio 2.5 was evaluated on music only; Stable Audio Open on sound effects only. The Small column is small-music for music rows and small-sfx for sound-effect rows.

  • Stable Audio 3.0 paper, Tables 3–5

    Stability's own results. FAD (lower is better) and CLAP (higher is better) use LAION-CLAP embeddings. OVL (overall quality), REL (text relevance) and MUS (musicality) are mean opinion scores from 14 listeners. Time is wall-clock seconds on an H200. Stable Audio 2.5 was evaluated on music only.

Only each model's own maker's number is shown; scores a maker reports for rival models are left out.