Robustness of Speech Quality Models to Pause Manipulation: Implications for Data Preprocessing
* Presenting author
Abstract:
Instrumental speech quality models use fixed-length windows that may contain varying amounts of silence. This study examines how pause manipulation affects model predictions and human perception. We evaluate two speech quality models, SQ-AST and NISQA, on 500 recordings under four pause conditions (2,000 stimuli total). A subset of 50 speakers was rated by 16 listeners for MOS, warmth, and competence. Group-level model predictions were stable (Delta MOS <0.1), but complete pause removal shifted more than 20% of speakers by over 0.3 MOS points. Human ratings were unaffected across all conditions and dimensions (p>.5). This justified pooling ratings across conditions, increasing the effective rater count from 4 to 16 per speaker. Against this ground-truth reference, moderate trimming improved SQ-AST (PCC: +6.6%; RMSE*: -7.4%), whereas complete pause removal harmed NISQA (PCC: -11.8%). The optimal preprocessing strategy is architecture-dependent, and the disconnect between model sensitivity and human insensitivity underscores the need for perceptual validation of preprocessing decisions.