Research digest — 2026-08-18¶
Automated sweep of the last 45 days, generated by scripts/research-watch.py. Titles and abstracts are arXiv's; anything under So what is a maintainer's note, not the authors' claim.
No PDFs are hosted here
Every entry links to the publisher's copy. We cite work; we do not redistribute it.
1 new across 1 topics.
dysarthric & atypical speech¶
Analyzing Speech Condition Effects in Dysarthric ASR: A Layer-wise Probing Study¶
Darwin Jelestin Muthu, Navya Gupta, Wei Lin Tay, Zhengchen Zhang, Daniel Wang Zhengkui, Rong Tong et al. — 2026-08-03
Automatic speech recognition (ASR) performance degrades sharply on dysarthric speech, yet how disordered articulation reshapes a model's internal representations is underexplored. We conduct a layer-wise probing analysis of a transformer ASR encoder on Mandarin dysarthric speech under three transcript-matched conditions: original dysarthric speech, speaker conditioned zero-shot TTS resynthesis, and unconditioned TTS. Probing reveals a task- and condition-dependent representation hierarchy: phoneme boundary information remains weak across all layers for dysarthric speech; phoneme identity is re…
So what: Diagnostic, not a remedy — and that is what makes it useful here. It probes where a transformer encoder loses dysarthric speech, and finds phoneme boundary information weak at every layer while phoneme identity survives further up. If that holds, it bounds what YazSes's decoder-side levers can do: initial_prompt vocabulary and the disfluency filter operate on tokens that are already downstream of the loss, so they can fix spelling of a recognised word but not a boundary the encoder never represented. It argues that atypical-speech support is an acoustic-side problem (padding, VAD, adapters) rather than a post-processing one — which is where accessibility.pre_speech_padding_ms and the dysfluency-friendly mode already sit, and it suggests the ceiling for that approach.
⚠ Mandarin, and a probing study, not an intervention. Transfer to English is unproven and no WER improvement is claimed, so this belongs in the reasoning behind a design, not in a claim to a user.