ANCHOR model cuts speech quality errors by 48% on partial audio

Assessing speech quality on partial audio has been a challenge for streaming and generative AI systems — traditional predictors assume full context and degrade on truncated inputs. To solve this, Zhuoyan Tao, Jiatong Shi, Hye-jin Shim, and Shinji Watanabe from Carnegie Mellon University and other institutions proposed ANCHOR (Autoregressive Non-intrusive Chunk-Ordered Refinement). The model extends prior work (ARECHO) by reformulating incremental quality assessment as a multi-resolution autoregressive task. ANCHOR employs dual-resolution tokens and a resolution-aware hierarchy, enabling coarse-to-fine refinement at both chunk and utterance levels within a single decoder. This design allows the model to produce accurate quality scores even when only the first few seconds of speech are available.

Experimental results demonstrate substantial robustness under partial input constraints. ANCHOR achieved a 48% reduction in PLCMOS error on 2-second prefixes compared to baseline methods. Convergence analysis identified an effective perceptual context horizon of 4–6 seconds — meaning the model relies on roughly that much audio to stabilize its quality estimate. A stress test with localized corruption also structured extrapolation biases, highlighting how perceptual quality accumulates over time. Accepted at Interspeech 2026, ANCHOR provides a practical framework for real-time quality monitoring in voice assistants, call centers, and generative audio pipelines.