Objectively Quantifying Speech Outcomes of Children with Cleft Palate

鼻音的知觉评估被认为是评估语音的一个关键组成部分

Perceptual assessment of hypernasality is considered a critical component when evaluating the speech of children with cleft lip and/or palate (CLP). However, most speech-language pathologists (SLPs) do not receive formal training for perceptual evaluation of speech and, as a result, research shows that the subjective ratings are inherently biased to the perceiver and exhibit considerable variability. In this project, we aim to develop an artificial intelligence (AI) algorithm that automatically evaluates speech along four dimensions deemed to be critically important by the Americleft Speech Outcomes Group (ASOG), namely speech acceptability, articulation, hypernasality, and audible nasal emissions. The AI algorithm in this project is based on an existing database of speech collected as a part of an NIH-funded project to develop reliable speech outcomes by improving the reliability of perceptual ratings by training clinicians (NIDCR DE019-01235, PI: Kathy Chapman). This database contains speech samples from 125 5-7 year olds along with multiple perceptual rating for each speech sample. The clinicians participating in this study were successfully trained using a new protocol from the Americleft Speech Outcomes Group and they exhibit excellent inter-clinician reliability. In SA1 we will develop an AI algorithm that automatically learns the relationship between a comprehensive set of speech acoustics and the average of the ASOG-trained expert ratings for each of the four perceptual dimensions. This approach is based on technology that the PIs have successfully used to evaluate dysarthric speech. Unique to these algorithms is modeling of perceptual judgments of trained experts using tools from statistical signal processing and AI. The output of the algorithms will map to a clinically- relevant scale, rather than to norm-referenced values that may or may not be meaningful. In SA2, we will evaluate the tool on new data by collecting new speech samples using a mobile app at a partner clinic using the same protocol as in the original study. Every collected sample will be further evaluated by ASOG trained clinicians. We will use this data to evaluate the accuracy of the AI model by comparing the model's predictions with the average of ASOG-trained experts. Preliminary results show promise that the proposed approach will yield a successful tool for accurately characterizing perceptual dimensions in the speech of children with CLP. These results indicate that a number of acoustic features that have been developed previously by the PIs accurately capture differences in hypernasality and articulation between the speech of three children with CLP (with varying severity). Furthermore, we show the success of our approach on a different, but related, task: objective evaluation of dysarthric speech. We show that an algorithm that automatically rates hypernasality performs on par with the judgment of human evaluators. The results of the proposed research will form the basis for a subsequent R01 proposal for the development and evaluation of a clinical tool to objectively quantify and track speech production in children with CLP.