Neural Codec Speech Quality Metrics Evaluated: scoreq and utmos Lead
A recent study assessed 45 objective speech-quality metrics tailored for neural audio codecs, testing 17 codec conditions. The findings revealed that the metrics demonstrated strong correlations with subjective listening scores for clear speech, with scoreq and utmos attaining the highest Pearson correlations. However, it was noted that non-intrusive metrics tend to saturate at elevated subjective quality levels. This research evaluates the reliability of these metrics for neural codecs and has been submitted on September 30, 2025. The paper is available on arXiv under ID 2509.24457.
Key facts
- Study evaluates 45 objective speech-quality metrics for neural audio codecs.
- Metrics correlated with subjective listening scores for clean speech.
- 17 codec conditions were tested.
- scoreq and utmos achieved the highest Pearson correlations.
- Non-intrusive metrics saturate at high subjective quality levels.
- Research addresses reliability of metrics for neural codecs.
- Paper submitted on September 30, 2025.
- Available on arXiv with ID 2509.24457.
Entities
Institutions
- arXiv