Bernard Muller

Research profile

Research summary

Bernard Muller is an early-stage researcher in clinical speech technology, profiled from public bibliographic data linked to ORCID 0009-0004-7494-305X; no institutional affiliation or country is confirmed in the available records. Two first-author arXiv preprints, both dated 2026 and reviewed by Evolve, define the visible footprint, and both concern dysarthric speech. The first introduces per-aetiology contrastive severity embeddings with phonological pseudo-labelling for multilingual dysarthric speech, reporting that aetiology-specific training beats pooled models for CP, PD and ALS severity classification by +23–40% relative F1; Evolve notes well-controlled but single-run experiments, SAP-dominated tests, and weak pseudo-label agreement (κ≈0.07). The second is a single-speaker case study comparing parameter-efficient fine-tuning variants (LoRA, DoRA, QLoRA and others) for per-patient dysarthric ASR across two ASR bases, where LoRA ties DoRA and outperforms 4-bit QLoRA, though n=1 limits generality. The Evolve reproducibility score is 55 (medium readiness on both papers), the collaboration score is 28 with 3 distinct co-authors, and the focus score is 0, reflecting an even split between two near-identical clinical speech field labels. With output, impact and activity axes lacking underlying data, this is a coherent but early profile: a methodical contributor to dysarthria-focused speech AI, centred on aetiology-aware severity modelling and parameter-efficient adaptation.

Research keywords

Highlights

• Per-aetiology training lifts dysarthria severity classification

  In a 2026 first-author arXiv preprint, Muller introduces contrastive severity embeddings with phonological pseudo-labelling for multilingual dysarthric speech; Evolve's review reports per-aetiology training beating pooled models for CP, PD and ALS severity by +23–40% relative F1, while flagging single-run experiments, SAP-dominated tests and weak pseudo-label agreement (κ≈0.07). Reproducibility rated medium.

• LoRA ties DoRA for per-patient dysarthric ASR in single-speaker study

  A 2026 first-author case study benchmarks parameter-efficient fine-tuning variants for per-patient dysarthric ASR on two ASR bases: LoRA ties DoRA and outperforms 4-bit QLoRA plus four other PEFT variants on one dysarthric speaker. Evolve calls the work careful but notes n=1 limits generality; reproducibility rated medium.

• A coherent niche: aetiology-aware and patient-specific clinical speech methods

  Taken together, the two 2026 preprints form one research line — making clinical speech models sensitive to who is speaking (aetiology, individual patient) rather than pooling all dysarthric speech — spanning HuBERT-based severity embeddings and PEFT-based ASR adaptation.

Timeline

1. 2026First-author arXiv preprint 'Choosing a PEFT Variant for Per-Patient Dysarthric ASR: A Single-Speaker Case Study on Two ASR Bases' appears in the Evolve catalog.
2. 2026First-author arXiv preprint 'Per-Aetiology Contrastive Severity Embeddings with Phonological Pseudo-Labelling for Multilingual Dysarthric Speech' is reviewed by Evolve, reporting +23–40% relative F1 over pooled models for CP, PD and ALS severity.
3. 2026Both catalog entries are dated 2026, with 3 distinct co-authors across the two papers, establishing an early two-paper footprint in clinical dysarthric speech technology.

Papers in the Evolve catalog

• 2026-09-18first author

  Per-Aetiology Contrastive Severity Embeddings with Phonological Pseudo-Labelling for Multilingual Dysarthric Speech

• 2026-09-02first author

  Choosing a PEFT Variant for Per-Patient Dysarthric ASR: A Single-Speaker Case Study on Two ASR Bases

Frequent co-authors