A cheap synthetic-data testbed reveals that connecting quantization noise to the computational graph via its standard deviation stabilizes straight-through estimator training, improving the descript-audio-codec without a commitment loss.
Title resolution pending
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
-
Efficient Evaluation of Quantization-Effects in Neural Codecs
A cheap synthetic-data testbed reveals that connecting quantization noise to the computational graph via its standard deviation stabilizes straight-through estimator training, improving the descript-audio-codec without a commitment loss.