News

Happy to share this work! When it comes to help, the two sides don’t value it the same way. Helpers tend to think it costs them little, while the people they helped see it as a bigger effort and a bigger favour. We found this asymmetry in our human participants. However, most large language models we tested showed a much weaker version of it. More details in our newly published paper in Frontiers in Psychology: https://doi.org/10.3389/fpsyg.2026.1831256

Different tones of voice (prosody) shape how we perceive a voice and make social judgments about the speaker. Does the same hold for AI-generated speech, which can be just as prosodically varied? Yes and no. In our latest publication in Computers in Human Behavior Reports, listeners heard 320 utterances crossing two prosodic styles (confident vs. doubtful) with two voice sources (human vs. AI-cloned), and rated 16 impressions such as humanlikeness, animateness, emotional richness, and attractiveness. We found that who's talking matters! Even without being told which voices were AI, listeners implicitly picked up on the source: confident prosody boosted impressions of AI voices too, but it didn't buy them the same social appeal it does for human voices — the perceptual mechanism doesn't transfer 100%. More details in this link.

On May 28 and 29, I was thrilled to attend CRBLM's Research Day and the Symposium on Computational and Mechanistic Analyses of Communicative Interactions. I was especially impressed to learn that some participants had driven for many hours from the United States to attend these two events because they considered them so worthwhile. Experiences like these remind me how fortunate I am to pursue my PhD at McGill University and to be part of such an intellectually stimulating research community.

"Who's talking" has long fascinated researchers. Our latest work, just published online in Neuropsychologia, shows that voice identity in human voices and their AI-cloned counterparts is processed in remarkably similar ways, as reflected in ERP responses to trained familiar vs. unfamiliar talkers. We also find that learning a speaker's name and later recognizing their voice elicits ERP patterns resembling those for intimately familiar people (e.g., family members), echoing Plante-Hébert et al. (2021). This points to a neat social trick: knowing someone's name can make a real difference when you're building a relationship with them!

AI voices are not stereotypically monotone, and neither are human voices. We introduced a pipeline to prepare and validate expressive human and AI voices that share comparable speaker identity. We also found that listeners are sensitive to AI-generated speech and do not readily accept that an AI voice shares the same identity as its human source, especially when speech prosody complicates the cognitive judgment. Our latest work, published in Computers in Human Behavior: Artificial Humans, is now available online.

I'm incredibly grateful to have passed my PhD Comprehensive Examination!

Major Paper:
How Do We Form Social Impressions of a Talker from Human and AI Speech? A Multilayered Framework
Committee: Nicole Li-Jessen and Nadine Lavan

Minor Paper:
Compensatory Adaptation in Multisensory Speech Perception but Persistent Lower-Order Multisensory Integration Impairments in Autism Spectrum Disorder: A Narrative Review
Committee: Linda Polka and Janet Bang

It was a great pleasure to work with Prof. Jiang to present a poster at Leiden University for SP2024. Most importantly, we did a lot of interesting work to bid for hosting Speech Prosody 2026 in Shanghai. Although the host eventually went to Philadelphia (Source), I found the experience exciting. We prepared a nice talk and organized a lucky draw event afterwards. We thank everyone we met at the conference!