Abstract
Speech corpora are collections of textual data derived from human verbal output and speech signals that can be processed from a variety of perspectives, including formal or semantic content, to serve analyses of different levels of linguistic organisation (phonemic, morphosyntactic, lexico-semantic and content information, prosody and intonation) and to serve analyses of important phenomena such as speech fluency and errors (non-fluencies). We focus on transcribing speech along with non-fluencies or dysfluencies, the detection of which plays an important role in the diagnosis of primary progressive aphasia, where we specifically examine articulation-based dysfluencies in nfvPPA speech. In this work, we propose SSDM 2.0, which is built on top of the current state-of-the-art system of dysfluency detection [1] and tackles its shortcomings via four main contributions: (1) We propose a novel Neural Articulatory Flow for deriving highly scalable, dysfluency-aware speech representations. (2) We develop a full-stack connectionist subsequence aligner to capture all major dysfluency types. (3) We introduce a mispronunciation prompt pipeline and consistency learning into LLMs to enable in-context dysfluency learning. (4) We curate and open-source Libri-Co-Dys [1], the largest co-dysfluency corpus to date. (5) We also present SSDM-L, a modular, non-end-to-end, lightweight model designed for clinical deployment. In clinical experiments on pathological speech transcription, we tested SSDM 2.0 using nfvPPA corpus primarily characterized by articulatory dysfluencies. Overall, SSDM 2.0 outperforms SSDM and all other dysfluency transcription models by a large margin. See our project demo page at https://berkeley-speech-group.github.io/SSDM2.0/.