Routing Sparse Autoencoder
A speech representation disentanglement method combining a TopK sparse autoencoder, routing-exclusive supervision, and cross-factor adversarial mechanisms. Operating on a frozen pre-trained encoder, it separates linguistic information from paralinguistic information into independent representation channels.
Timeline (last 90 days)
The study proposes routing linguistic and paralinguistic information into different representation channels to achieve controllable disentanglement without retraining the encoder.
The method splits the encoder output features into two routes: a linguistic route and a paralinguistic route.
Experiments using independent probes showed that linguistic information is stronger in the linguistic route and attenuated in the paralinguistic route.
Paralinguistic factors (speaker identity, emotion, prosody) are preserved in the paralinguistic route and significantly reduced in the linguistic route.
Experiments swapping the factors carried by a given route in the feature space showed that the swapped factors were successfully transferred while the unmodified route information was largely preserved.
The method validated the consistency of route-selective separation across multiple dimensions, including cross-encoder, cross-corpus, cross-independent-probe, and cross-representation-level intervention.
All Facts (6)
The study proposes routing linguistic and paralinguistic information into different representation channels to achieve controllable disentanglement without retraining the encoder.
50%UnverifiedThe method splits the encoder output features into two routes: a linguistic route and a paralinguistic route.
50%UnverifiedExperiments using independent probes showed that linguistic information is stronger in the linguistic route and attenuated in the paralinguistic route.
50%UnverifiedParalinguistic factors (speaker identity, emotion, prosody) are preserved in the paralinguistic route and significantly reduced in the linguistic route.
50%UnverifiedExperiments swapping the factors carried by a given route in the feature space showed that the swapped factors were successfully transferred while the unmodified route information was largely preserved.
50%UnverifiedThe method validated the consistency of route-selective separation across multiple dimensions, including cross-encoder, cross-corpus, cross-independent-probe, and cross-representation-level intervention.
50%