I will be presenting the research and results from my PhD thesis entitled “Data-driven Speech Intelligibility Prediction.” Speech Intelligibility (SI) is a number between zero and one that relates to a speech signal, for example a few minutes of a person talking in a noisy environment. SI describes the proportion of the words in this signal, that are recognized and understood by listeners on average. SI is of interest to industries and researchers that work with technologies or devices that process speech, such as hearing aids, headsets, telephony, or other communications technologies.

We measure SI by carrying out a so-called listening test, in which a panel of human listeners are asked to identify the words in speech signals presented to them. This is a time-consuming process, which is why algorithmic methods, designed to predict SI without the need for human listeners, are desired.

My research focuses on predicting SI using machine learning models and algorithms, popularly known under the umbrella term artificial intelligence. These models have been exposed to data and results from collections of listening tests, and thereby learned how to predict the SI of speech signals in various circumstances, e.g., certain types of background noise and digital processing.

My research shows that data-driven SI predictors are more accurate than existing alternatives under circumstances they have seen in their training. Unfortunately, they do not work as well in circumstances they have not been trained for, e.g., a new type of background noise. This problem is caused by the fact that focused listening test data is currently too scarce to train data-driven SI predictors effectively. Part of my research has focused on increasing the accuracy of data-driven SI predictors when applying them under new circumstances.

The presentation will include an introduction to the field of data-driven SI prediction, as well as the motivations, methods, and results of my research in this field.