Presented By: Department of Linguistics
PhonDi Discussion Group
Will Styler, Modeling Human Speech Perception Using Machine Learning
Will Styler will be discussion "Modeling Human Speech Perception Using Machine Learning."
Abstract
One struggle in identifying the acoustic cues used in speech perception is the near infinite number of possible features usable by humans. Here we describe a more efficient, machine-learning-based alternative. Acoustic measurements of 29 features were used to train a Support Vector Machine, allowing the classification of English vowels as “oral” or “nasalized”. The best-performing features were then tested using modified experimental stimuli with human listeners. The SVM model and human listeners showed similar patterns of confusion and perception, suggesting that SVMs can be used to predict the utility of different features for human perception.
Abstract
One struggle in identifying the acoustic cues used in speech perception is the near infinite number of possible features usable by humans. Here we describe a more efficient, machine-learning-based alternative. Acoustic measurements of 29 features were used to train a Support Vector Machine, allowing the classification of English vowels as “oral” or “nasalized”. The best-performing features were then tested using modified experimental stimuli with human listeners. The SVM model and human listeners showed similar patterns of confusion and perception, suggesting that SVMs can be used to predict the utility of different features for human perception.
Explore Similar Events
-
Loading Similar Events...