A layer-wise correlation-permutation alignment lets a merged HuBERT-MERT encoder keep speech performance and improve music scores, beating linear interpolation by 19.5 points on the authors' aggregate score.
Language models are super mario: Absorbing abilities from homologous models as a free lunch,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.SD 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
A correlation-permutation approach for speech-music encoders model merging
A layer-wise correlation-permutation alignment lets a merged HuBERT-MERT encoder keep speech performance and improve music scores, beating linear interpolation by 19.5 points on the authors' aggregate score.