Calibration-Reasoning Framework for Descriptive Speech Quality Assessment

A team from EPFL, led by Elizaveta Kostenok, Mathieu Salzmann, and Milos Cernak, has published a new paper detailing a 'Calibration-Reasoning Framework' designed to revolutionize how AI assesses speech quality. The core innovation is moving beyond a single, opaque Mean Opinion Score (MOS) to provide multidimensional, explainable analysis. The framework first calibrates a foundational Audio Large Language Model to understand predefined perceptual dimensions (e.g., noisiness, distortion). It then uses a novel reinforcement learning technique called Group Relative Policy Optimization (GRPO) with dimension-specific rewards to dramatically enhance the model's descriptive accuracy and its ability to temporally localize—or pinpoint in time—specific audio issues.

This two-stage post-training method achieved state-of-the-art results, including a 0.71 mean Pearson Correlation Coefficient (PCC) on the multidimensional QualiSpeech benchmark. Crucially, the reinforcement learning-based reasoning stage drove a 13% improvement in MOS prediction accuracy compared to baseline methods. The fine-grained GRPO rewards are key, enabling the model not just to detect that audio quality is poor, but to explain *why* by classifying artifacts (like 'click at 2.3 seconds' or 'background hum between 5-7 seconds') with unprecedented temporal precision. The work has been submitted for presentation at Interspeech 2026.