Work

Speech AI that works on children's real-world audio Case study →

Built the multi-country data consortium and annotation pipeline behind the models; co-led VTC; co-supervised BabyHuBERT · 2019–2026

Open-source models that tell who is speaking, and when, in day-long recordings worn by young children: state of the art across 40+ languages, approaching human-annotator agreement.

Making sensitive child audio usable, legally and ethically Case study →

Led data-protection practice and model governance; co-authored a 28-country handbook · 2021–2024

Turned legal and ethical uncertainty into working processes: a 28-country data-protection handbook, risk assessments, an ethics board, and a licensing approach for releasing a model trained on children's audio.

Recording hardware that works from Paris to Vanuatu Case study →

Led an international product-development initiative · 2017–2024

Defined how to choose and attach child-worn recorders safely across cultures, climates and budgets. Piloted on four continents.

Speaker diarization benchmarks: DIHARD I & II and JSALT 2019

Co-organizer · 2018–2019

Co-organized international challenges that set the benchmark for 'who spoke when' in hard, real-world audio, and the speaker diarization track at the JSALT 2019 workshop.

DIHARD I challenge site ↗DIHARD II publication ↗JSALT 2019 workshop team ↗JSALT 2019 paper ↗

Annotation pipelines for sensitive audio, including crowdsourcing

Designed taxonomies, training and quality control · 2018–2026

Built annotation categories, annotator training and quality control (inter-annotator reliability and derived metrics) for human and machine labels, and adapted crowdsourcing to recordings that cannot be widely shared.

Representative training and benchmark datasets

Supervised dataset creation and splitting protocols · 2019–2026

Supervised multi-level acoustic and text datasets for training and benchmarking models, with stratified sampling to audit demographic and acoustic bias, and grouped splits that prevent data leakage between training and test sets.

Evidence synthesis: meta-analysis and a hybrid meta/mega-analytic method

Lead analyst · 2014–2021

Ran systematic reviews and meta-analyses end to end, including bias audits at study and population level, and developed a hybrid meta- and mega-analytic method for handling missing data.

Measures of child development that are fair across cultures and languages

Researcher · 2009–2026

Developed measures of cognitive and language development that work fairly across cultures, languages and multilingual settings, building on fieldwork in Bolivia and Papua New Guinea and advising teams in low- and middle-income countries, including Timor-Leste, Solomon Islands, India and Ghana.

Teaching: AI for Social Good and Cognitive Science in the Real World

Co-creator and instructor · 2021–2026

Co-created and taught interdisciplinary courses on the societal implications of research and AI.