SPEAR
A self-supervised speech pre-training encoder trained on large-scale unlabeled speech data. It generates universal speech representations that integrate both linguistic and paralinguistic information, supporting a range of downstream speech understanding applications.
Timeline (last 90 days)
Self-supervised speech encoders compress linguistic and paralinguistic information into the same shared, entangled representation space.
The entire process is based on a frozen encoder, leaving the original weights of SPEAR and WavLM unchanged, and only reorganizing representations at the output level.
All Facts (2)
Self-supervised speech encoders compress linguistic and paralinguistic information into the same shared, entangled representation space.
50%UnverifiedThe entire process is based on a frozen encoder, leaving the original weights of SPEAR and WavLM unchanged, and only reorganizing representations at the output level.
50%