2020/09/23 by Vikram C. Mathad, Mathad, Vikram C., Nancy J. Scherer +7 · 2 citations
Biochemistry, Genetics and Molecular Biology · #Audio and Speech Processing (eess.AS) #Cleft Lip and Palate Research #FOS: Electrical engineering #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2009.11354
openalex publication_date 2020/09/23 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Objectives: Evaluation of hypernasality requires extensive perceptual\ntraining by clinicians and extending this training on a large scale\ninternationally is untenable; this compounds the health disparities that\nalready exist among children with cleft. In this work, we present the objective\nhypernasality measure (OHM), a speech analytics algorithm that automatically\nmeasures hypernasality in speech, and validate it relative to a group of\ntrained clinicians. Methods: We trained a deep neural network (DNN) on\napproximately 100 hours of a publicly-available healthy speech corpus to detect\nthe presence of nasal acoustic cues generated through the production of nasal\nconsonants and nasalized phonemes in speech. Importantly, this model does not\nrequire any clinical data for training. The posterior probabilities of the deep\nlearning model were aggregated at the sentence and speaker-levels to compute\nthe OHM.\n Results: The results showed that the OHM was significantly correlated with\nthe perceptual hypernasality ratings in the Americleft database ( r=0.797,\n~p<0.001), and with the New Mexico Cleft Palate Center (NMCPC) database\n(r=0.713,p<0.001). In addition, we evaluated the relationship between the OHM\nand articulation errors; the sensitivity of the OHM in detecting the presence\nof very mild hypernasality; and establishing the internal reliability of the\nmetric. Further, the performance of OHM was compared with a DNN regression\nalgorithm directly trained on the hypernasal speech samples. Significance: The\nresults indicate that the OHM is able to rate the severity of hypernasality on\npar with Americleft-trained clinicians on this dataset.\n