Problem
Research questions and friction points this paper is trying to address.
Speech Quality Assessment
Mean Opinion Score
Domain Generalization
Label-Free
ProxyMOS: Label-Free Speech Quality Assessment by Multi-Teacher Distillation with Adaptive Routing
🤖 AI Summary
This study addresses the high cost of manual annotation and the limited cross-domain generalization of non-intrusive models in speech quality assessment by proposing an unsupervised evaluation framework based on multi-teacher knowledge distillation. Utilizing wav2vec 2.0 as the backbone, the method introduces an adaptive routing ensemble mechanism to generate high-quality pseudo-labels for training a student model. Furthermore, subset search is employed to optimize predictor combinations, ensuring that incorporating weaker predictors does not degrade performance. Experimental results demonstrate that the proposed model achieves Spearman correlation coefficients of 0.802 and 0.636 on the URGENT and MOS260 benchmarks, respectively. Notably, it surpasses the best teacher model with only a single forward inference pass, thereby enabling efficient and robust cross-domain speech quality assessment.
📝 Abstract
Human mean opinion scores (MOS) are costly to collect, and non-intrusive MOS predictors degrade sharply outside their training domain. ProxyMOS turns a pool of public MOS predictors into a single stronger model without new human labels. Eight predictors are benchmarked against human ratings; the five most informative enter a subset search under uniform, correlation-weighted, error-weighted, MSE-optimised and adaptive per-utterance routing; and the best routed four-model ensemble labels 807k unlabeled utterances that train a wav2vec 2.0 student. On URGENT the student reaches Spearman $ρ=0.802$ against $0.773$ for the best teacher. On mos260, a new Russian TTS benchmark of 4,600 utterances from 38 synthesis conditions, it reaches $ρ=0.636$ against $0.613$ per utterance and $0.95$ per condition, matching its own routed ensemble in one forward pass. Adaptive routing is the only rule that does not degrade when weak predictors are added. Model, ONNX exports and mos260 are released. It's about 950 characters; arXiv's limit is 1,920. I kept $ρ$ because arXiv renders it on the abstract page. If you'd rather avoid math, replace $ρ=0.802$ with rho = 0.802 and do the same for the other $...$ values.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Label-Free Speech Quality Assessment
Multi-Teacher Distillation
Adaptive Routing
Pseudo-Labeling
wav2vec 2.0
🔎 Similar Papers
No similar papers found.
