Fine-tuning WavLM on the diarization task before structured pruning yields 80% parameter removal at nearly unchanged diarization error, with 2.6x to 4x faster GPU inference.
For Transformer layers, the 9th and 10th are almost completely removed after pruning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
eess.AS 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Fine-tune Before Structured Pruning: Towards Compact and Accurate Self-Supervised Models for Speaker Diarization
Fine-tuning WavLM on the diarization task before structured pruning yields 80% parameter removal at nearly unchanged diarization error, with 2.6x to 4x faster GPU inference.