MOS-Naturalness Scale

The MOS-Naturalness score estimated via NISQA-TTS (Non-Intrusive Speech Quality Assessment - Text-To-Speech) represents the perceived naturalness and human-likeness of synthetic speech. It follows the ITU-T P.85 absolute category rating (ACR) protocol, where listeners judge samples on a 1–5 scale:

Practical Range and Expected Scores

Although the theoretical scale runs from 1 to 5, in real listening tests the effective range is narrower:

• Human reference recordings typically score ≈ 4.5 – 4.8, rarely a perfect 5.0.
• Modern neural TTS systems (e.g., Tacotron 2, VITS, Vall-E) achieve ≈ 3.9 – 4.4.
• Mid-quality synthetic voices cluster around 2.5 – 3.5.
• Clearly robotic or degraded speech falls below 2.0.

NISQA-TTS reproduces this empirical distribution: even for flawless natural recordings, predicted values seldom exceed 4.6 – 4.7, while highly natural TTS outputs occupy the 3.5 – 4.3 range.

Why the Ceiling Is Lower Than 5.0

This compression of the scale is not a modeling flaw but a human-rating phenomenon known as central tendency bias. When people use bounded scales (like the 1–5 MOS), they tend to avoid extreme options — hesitating to label anything “perfect” or “terrible.” The result is a clustering of ratings toward the middle and an empirical ceiling below 5.0. Consequently, MOS values around 4.5 already correspond to perfectly natural human speech in subjective perception.

Summary

The MOS-Naturalness scale in NISQA-TTS quantifies how natural or human-like speech sounds on a 1–5 scale aligned with ITU-T P.85. In practice, due to listener behavior and dataset calibration, the usable range is approximately 1.0 – 4.7, with 4.5 – 4.8 representing human recordings and around 4.0 indicating very natural synthetic speech. Scores should therefore be interpreted relatively—not as absolute percentages of perfection, but as perceptually grounded indicators of naturalness.