Kate Knill

Research profile

Research summary

Kate Knill is a speech-technology researcher whose profile is name-resolved from public OpenAlex bibliographic data; no institution, country, or ORCID is attached to this record. The available evidence — two 2026 arXiv papers reviewed in the Evolve catalog plus a sparse OpenAlex snapshot — points to a concentrated research program in low-resource automatic speech recognition built on parameter-efficient fine-tuning (PEFT) of large speech models such as Whisper and Qwen3. In the more striking result, stacking a frozen related-language adapter beneath a target-language adapter outperforms full fine-tuning by 5–8% WER on three unseen languages; the companion work applies a Fisher-whitened cross-covariance objective with adaptive rank allocation, yielding small, partly significant gains. In both papers Knill sits in the last author position without corresponding-author designation, consistent with a senior supervisory role on a small team of five distinct co-authors recorded, including Mengjie Qian, Thai Thi Thanh Thao Dang, Asmee Mishra, and Brechtje Post. The OpenAlex topics — speech recognition and synthesis, phonetics and phonology, and second language acquisition — suggest the engineering work sits inside a broader speech-and-language agenda. Bibliometric coverage is thin: one indexed work (2024), h-index 0, i10-index 0, and one work in the last three years, so output and impact remain unscored and the reviewed findings await peer review.

Research keywords

Highlights

• Frozen related-language adapter stacking cuts WER by 5–8%

  In “Sequential Adapter Stacking for Cross-Lingual Low-Resource ASR” (arXiv, 2026), placing a frozen adapter trained on a related language beneath a target-language adapter beats full fine-tuning by 5–8% WER on three unseen languages, in a controlled single-backbone setup; Knill authors in the last position.

• Fisher-whitened, adaptive-rank PEFT for low-resource ASR

  In “Adaptive Fisher-Whitened Cross-Covariance for Low-Resource Speech Recognition” (arXiv, 2026), a task-informed FCCA parameter-efficient fine-tuning scheme with adaptive rank allocation is evaluated on Whisper and Qwen3; the adaptive-rank variant gives small, partly significant gains — credible but single-seed evidence per the Evolve review.

• A speech-and-language program bridging engineering and phonetics

  OpenAlex topics pair Speech Recognition and Synthesis with Phonetics and Phonology Research and Second Language Acquisition and Learning, suggesting the ASR engineering sits inside a broader agenda spanning phonetics and second-language learning; all indexed work maps to a single field, speech processing (ASR).

Timeline

1. 2024OpenAlex records its only indexed work for this researcher in this year — the sole entry in the bibliometric coverage window and the anchor of the sparse activity signal.
2. 2026Co-authors “Sequential Adapter Stacking for Cross-Lingual Low-Resource ASR” (arXiv) in the last author position; stacking a frozen related-language adapter beats full fine-tuning by 5–8% WER on three unseen languages.
3. 2026Co-authors “Adaptive Fisher-Whitened Cross-Covariance for Low-Resource Speech Recognition” (arXiv) in the last author position; the adaptive-rank FCCA PEFT variant shows small, partly significant gains on Whisper and Qwen3.

Papers in the Evolve catalog

• 2026-09-24

  Adaptive Fisher-Whitened Cross-Covariance for Low-Resource Speech Recognition

• 2026-09-14

  Sequential Adapter Stacking for Cross-Lingual Low-Resource ASR

Frequent co-authors