EPJ B Highlight - Understanding speech with a new model of word recognition

Details

Published on 15 April 2025

Researchers found some surprising differences in the way humans handle long and short words

A new dynamical model of speech recognition has revealed the very different ways that humans perceive short and long words in everyday speech. The authors of the research published in EPJ B, Jean-Marc Luck of the Université Paris-Saclay and Anita Mehta, formerly of the Faculty of Linguistics, Oxford and currently at St Catherine’s College, University of Oxford, take a radically different approach to speech perception.

“Our emphasis lies less in getting exact answers to individual instances of word retrieval, which underlies the usual computational linguistics approaches, and more in understanding the global aspects of speech perception from a statistical physics point of view by examining the salient features of large collections of instances,” the authors said. “The present model builds on an earlier one by introducing correlations between neighbouring sounds, which exist naturally in world languages.”

The authors added that the resulting lexicon is rich in short words, and less so in longer ones, which is in agreement with word length distributions in most languages. The authors then constructed an algorithm that models the perception of these two-word categories in the presence of mishearings.

“Our findings were that short words are quickly retrieved, while the retrieval of longer words is somewhat slower, with a finite probability of getting lost altogether — exactly as one might expect in everyday life,” the authors said. “Many of the results of this work were surprisingly true to life, especially to do with the flying Dutchman-like wandering of the algorithm representing the speech perception process in the brain for long, misheard words, in the vicinity of the ‘true’ word, without ever really settling on it. In the current model, this only occurs in a small proportion of cases of word retrieval.”

The team said that possible applications of these findings range from developing better tools for language laboratories to improving interventions for speech therapy.

“We are currently working on a model where the presence of successive mishearings leads to an extremely large probability that the algorithm slows down dramatically and typically fails completely to retrieve a long and complex word,” the authors concluded.