Speech Science Forum - UCL SHaPS 2nd year PhD students

Maya Hola

Phonetic Distance Predicts Intelligibility in Noise in Homogenous Accent Group

Phonetic distance - a quantitative measure of acoustic-phonetic divergence between speakers’ realizations of the same utterance - is known to affect how well speakers understand each other in noise, particularly across accents. Whether this effect extends to within-accent variation, where differences are comparatively subtle, remains unclear. We tested 40 Standard Southern British English adults on a sentence recognition task across four signal-to-noise ratio conditions. We measured phonetic distance between each talker and listener using four types of metrics, which I’ll briefly demonstrate during the talk. We also measured talker typicality within the SSBE group as the average distance from each talker to all other talkers. Talker typicality was a robust predictor of intelligibility, with pairwise talker–listener distance providing additional explanatory power. These findings extend accent-distance accounts to the idiolectal level, showing that fine-grained phonetic variation within a single accent impacts communication in noise.

– 

Yuhan Huang

How Source- and Filter-Related Acoustics Drive Automatic Speaker and Emotion Recognition

Profiling speakers through their voices has long been approached as both a human perceptual and an automatic recognition task, yet the acoustic dynamics underlying machine learning models remain largely opaque. Grounded in source-filter theory, this talk presents experiments I have conducted during my PhD, investigating the relative contributions of voice source and vocal tract filter to automatic speaker recognition (ASR) and emotion recognition (ER) using eGeMAPS features and an interpretable machine learning model. I will discuss findings from a modal and whispered speech paradigm across 60 actors, which largely dissociates source- from filter-related acoustic information, and share some interesting patterns that emerged from the results. I will also outline plans for my next experiments, which will extend this work to incorporate human perceptual data alongside the acoustic modelling. 

–

Xinyi Zhao

Learning variation in a second language: Effects of listener experience and attention control on regional accent processing

When speakers move to a new region, they often adjust their speech to fit the local community. Most research has focused on monolinguals, but recent studies show that exposure to regional accents can also improve second language (L2) learners’ awareness and identification of regional variation. However, it remains unclear whether such exposure also supports comprehension. It is also not well understood how factors such as length of residence (LoR), L2 proficiency, and attention control influence the ability to understand and recognise different L2 varieties. This talk presents the first study of my PhD, examining how experience and cognitive factors shape L2 listeners’ perception of regional accents.