Word-level intelligibility model for the third Clarity Prediction Challenge
Mark Huckvale
This paper presents a speech intelligibility model for the third Clarity Prediction challenge based on an analysis of word-level intelligibility in the training dataset. Using the given test prompts, a word-level alignment was performed on the reference audio, and this was then used to extract information from the test audio, including word-level measures of acoustic and phonetic distortion. Lexical properties of the words were also obtained using other language resources, including phone count, syllable count, word frequency, trigram frequency and number of lexical neighbours. We present an analysis showing how the intelligibility of individual words relates to these properties and build a classification model that uses them to predict word intelligibility. We show that sentence level intelligibility predictions derived from a word-level intelligibility prediction model gives better performance than a model based on whole sentences. On the evaluation data set, the model achieved a correlation of 0.759 and a RMS prediction error of 26.9%.