Welcome, Bienvenue !

I am a Machine Learning Researcher in Apple's Machine Learning Research team, led by Samy Bengio, and I am based in Copenhagen, Denmark.

My research focuses on how speech and language models learn, represent, and use linguistic information. I am particularly interested in multilingual and speech-language models, including how language information is represented and used, what models gain from operating directly on speech, and how training, modality, and multilinguality shape their behaviour. My current work spans multilingual speech and language modelling, multimodal learning, machine interpreting and translation, and model evaluation.

I completed my PhD in Machine Learning and Cognitive Science at the Ecole Normale Supérieure, under the supervision of Emmanuel Dupoux and Guillaume Wisniewski. During my PhD, I used unsupervised and self-supervised speech models to study early language acquisition, including phonetic and word learning and bilingual development. I was also an AI Research Scientist Intern at Meta. Before that, I earned an MSc in Speech and Language Processing from the University of Edinburgh, and a BSc in Psychology from City, University of London. I also worked as a speech recognition engineer before starting my PhD.

Research 

• Multilingual learning and language information

I study how multilingual models represent and use language information, how this interacts with other aspects of their representations, and when distinguishing between languages is useful. My work includes language discrimination, multilingual speech and text models, data quality and selection, and multimodal approaches to multilingual learning.

• Speech-language models and spoken communication

I study models that operate directly on speech and what they gain from the speech modality. This includes speech-language models, prosody and other speech-specific information, machine interpreting and speech translation, and how spoken and text-based language understanding differ.

• Evaluation and understanding of learned representations

I develop controlled evaluations to understand what information models encode and use, including phonetic, prosodic, language and semantic information. I am particularly interested in methods that isolate specific capabilities and reveal behaviour that aggregate task metrics can miss.

My approach is strongly influenced by cognitive science. I often treat models as systems to study as well as systems to improve, using controlled experiments to understand what they have learned and which aspects of their input they rely on. I am also interested in building closer connections between machine learning and psycholinguistics; you can find the slides from our Interspeech tutorial on the topic here.

I also supervise research interns and develop projects across these areas, including current work on multilingual speech and speech-language models.

Latest news

• 31/08/2026 - [🌍 TRAVEL] I’ll be travelling to Sydney for Interspeech 2026 at the end of September! If you’re attending, do check out our work “Which Data Matter? Embedding-Based Data Selection for Speech Recognition”, and feel free to reach out if you’d like to meet up!

• 04/05/2025 - [🌎 TRAVEL]   I’ll be at ICASSP 2026 in Barcelona! If you’re attending, do check out our paper "Leveraging Audio-Visual Data to Reduce the Multilingual Gap in Self-Supervised Speech Models", led by María Andrea Cruz Blandón during her internship with us at Apple. Happy to meet up if you’re around!

• 01/05/2026 - [👋 UPDATE] After several busy months on maternity leave, I’m happy to be back at work and looking forward to reconnecting with colleagues and getting back to research!

• 07/11/2025 - [🏆 AWARD]  We received an SAC Highlight Award for our paper "Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks" at EMNLP 2025 in Suzhou!

• 24/09/2025 - [🤞 PREPRINT]   New preprint out : "Leveraging Audio-Visual Data to Reduce the Multilingual Gap in Self-Supervised Speech Models".  In this work, led by María Andrea Cruz Blandón during her internship at Apple, we show that visual grounding helps reduce the multilingual gap present in self-supervised speech models. 

• 19/09/2025 - [📄 PAPERS]  We’ll be presenting three papers at EMNLP 2025! And as a highlight, our paper "Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks", was nominated for the the SAC Highlight Awards, fingers crossed!

• 16/08/2025 - [🎥 SLIDES]    The slides for our Interspeech tutorial on Language Acquisition and Speech Technology are available online : https://zenodo.org/records/17018214

• 16/08/2025 - [🌎 TRAVEL]    I’ll be in Rotterdam for Interspeech 2025, where I’m co-presenting a tutorial on Language Acquisition and Speech Technology (details below). 

• 27/07/2025 - [🌎 TRAVEL]    Heading to Vienna for ACL 202! ! I’ll be around for the conference and the IWSLT workshop, happy to meet up if you’re attending.

• 22/05/2025 - [🤞 PREPRINT]   New preprint out : "Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks".  We apply ABX tasks to multilingual text models to understand how form and meaning representations are organised.

• 22/04/2025 - [🧑‍🏫 TUTORIAL]    I will be presenting, along with Emmanuel Dupoux and Okko Räsänen, a tutorial at Interspeech 2025 on the topic : Speech Technology Meets Early Language Acquisition: How Interdisciplinary Efforts Benefit Both Fields. Looking forward to seeing many of you!

• 21/04/2025 - [📄 PAPER]   Our paper with Ansgar Endress, entitled "The specificity of sequential statistical learning: Statistical learning accumulates predictive information from unstructured input but is dissociable from (declarative) memory for words", was just released in Cognition. It features some of my very early work as an psychology undergraduate about 10 years ago, nice to see it out!

• 06/01/2025 - [📄 PAPER]   New paper out in Developmental Science :  Simulating Early Phonetic and Word Learning Without Linguistic Categories. This work, which is a significant part of both Marvin Lavechin and my PhD research, looks at simulating early language acquisition with SSL speech models, with a cognitive perspective.