wavalign

WavAlign introduces a modality-aware adaptive post-training method that uses constrained preference updates and explicit anchoring to boost both semantic quality and speech expressiveness in end-to-end spoken dialogue models.