Guest Talks in Spring Semester 2026

The Phonetics and Speech Sciences (PaSS) Lunch Meeting will host Prof. Carolyn McGettigan (University College London) and Dr Nadine Lavan (Queen Mary University of London) as guest speakers on Wednesday, 13 May 2026,  12:15–13:45 at BIN-O-K.11/12/13.

Speaker: Prof. Carolyn McGettigan (University College London)
Title: Investigating the perception and moral acceptability of voice identity cloning
Speaker: Dr Nadine Lavan (Queen Mary University of London).
Title: Forming first impressions from voices
Wednesday, 13 May 2026, 12:15–13:45
Room:  BIN-O-K.11/12/13 (please note the change of venue from the usual PaSS Lunch Meetings).

Abstract: Investigating the perception and moral acceptability of voice identity cloning

In this talk I’ll present highlights from my British Academy Mid-Career Fellowship (2023-24) in which we ran a series of studies investigating how human listeners deal with state-of-the-art voice identity clones. I’ll present some work that was directly focused on cloning, asking how listeners perceive familiar and unfamiliar clones compared with human speech recordings, and how they feel about the uses of cloning technology in everyday life. Some of our other work has instead harnessed voice cloning technology to ask different questions about human listening: here, I will mention our experiments investigating how narrator voices might impact audiobook listening experiences and outcomes.

Abstract: Forming first impressions from voices

As soon as we hear a voice, we very quickly form an impression of the person we are hearing: Are they young or old? Are they friendly or grumpy? Do they sound posh? Impressions are often quite detailed and include a whole range of different characteristics. While some aspects of these first impressions are accurate to some degree (e.g., rough age estimates, gender judgments), other aspects do not seem to have any link to the person's actual characteristics (e.g., personality-related judgments).

In this talk, I will trace how listeners put together these complex impressions from the first few milliseconds of hearing a voice to having formed a full first impression. I will also discuss factors, such as 'personal taste' and shared conceptual knowledge (or stereotypes), that can shape first impressions from voices.

We are pleased to welcome Dr. Joel Niklaus (Hugging Face) as a guest speaker in the course "Text Generation with Language Models" on 18 May 2026, 10:15–11:45 at AND-3-02.

Speaker: Dr. Joel Niklaus (Hugging Face)
Title: Scaling Synthetic Text Generation to Trillions of Tokens
Monday, 18th May 2026, 10:15–11:45
Room: AND-3-02, Andreasstrasse 15, 8050 Zürich

We are pleased to welcome Dr. Clara Meister (EPFL) as a guest speaker in the course "Modul Retrieval-Augmented Generation" on 22 May 2026, 10:15–11:45 at AFL-F-121.

Speaker: Dr. Clara Meister (EPFL)
Title: Multilingual NLP in the LLM Era
Friday, 22th May 2026, 10:15–11:45
Room:  AFL-F-121, Affolternstrasse 56, 8050 Zürich

Abstract

Modern language models are increasingly expected to work across hundreds of languages, yet there is a large gap between "supports N languages" and equitable cross-lingual utility. This lecture presents the standard recipe for training today's multilingual LMs and then revisits each step to expose where it breaks. Specifically, we will examine data quality and language identification for the long tail of low-resource languages, tokenization disparities that translate into user costs and performance inequities, and the curse of multilinguality that emerges from joint training with many languages. We will then look at how modern LLM adaptation methods work to mitigate some of these findings, and why current benchmarks systematically overestimate multilingual competence. Lastly, we will discuss some of the open problems at the forefront of multilingual LLM research, e.g., cultural and dialectal evaluation, code-switching, and tokenization fairness.

Bio

Clara Meister is a postdoctoral researcher in the NLP Lab at EPFL, where she works on tokenization and multilingual language modeling with a focus on low-resource settings. She completed her PhD at ETH Zürich, where her research foci were decoding algorithms for probabilistic language generators and information-theoretic analyses of human language processing. She continues to lecture in the MAS in AI and Digital Technology programme at ETH Zürich, and is a co-founder and organizer of Zürich AI, Switzerland's largest machine learning meetup.