Deep Learning Approaches to Assess Speech Intelligibility of Head and Neck Cancer

Loss of speech intelligibility is commonly found in the post-treatment of conditions that affect the vocal tract, such as head and neck cancer. Due to this, perceptual evaluations are still the most widely used method to clinically assess speech intelligibility. On the other hand, these evaluations are known to be highly subjective, biased and time-consuming since the evaluation can be conditioned by the practitioner, or patients previously assessed. In order to tackle these issues, an automatic assessment has been seen as a growing and viable alternative, that could provide more objective, faster and unbiased measures. In the present work, we explore distinct ways to predict speech intelligibility based on the different granularity levels of sentence, word and phoneme. The results from the proposed granular models suggest correlations with the perceptual intelligibility ranging from 0.80 to as high as 0.89 when applied to the French head and neck cancer speech corpus. The results also suggest a correlation up to 0.91 when merging all granular systems. Several conclusions are drawn from each granularity level, namely concerning specific types of words and phonemes that play different levels of relevance for the intelligibility of distinct speakers. Moreover, a study on the individual modelling of a set of perceptual judges is also presented. The study showcased that different judge profiles emerge from the perceptual and the automatic set of judges. Similarly to the granular systems, the results suggest that an automatic approach can indeed be seen as more uniform and objective. This leaves the possibility of these approaches being implemented in clinical environments to either serve as a second opinion or to free the practitioner to perform other relevant tasks.