Title:  Improved Automatic Analysis Methods for Lena-Obtained Audio Recordings of Children with Autism Spectrum Disorder

Committee: 

Dr. Anderson, Advisor   

Dr. Lee, Chair

Dr. Inan

Dr. Clements

Abstract:

The objective of the proposed research is to develop novel automatic analysis methods for audio recordings of children with autism spectrum disorder (ASD). The first area of exploration is speaker diarization, which allows us to identify ”who spoke when” and thus differentiate between the audio segments of a child vocalizing and those of a parent. Once diarization is completed, we can compute language behavior statistics that clinicians can use for diagnosis purposes and for tracking treatment effectiveness. Audio recordings of children with ASD are often recorded in a clinic setting or at home, and diarization performance is directly impacted by the environment where the audio is obtained. This work will thus center around the development of an environment-aware speaker diarization model. Traditional i-vector-based speaker modeling can be supplemented with environmental factors (e-vectors) for increased accuracy. Additional work on audio diarization will involve using time delay neural network (TDNN) techniques, namely x-vectors. Unlike i-vectors, which are trained generatively, x-vectors utilize speaker labels for direct supervision during training. X-vector modeling addresses the issue of limited training data by utilizing transfer learning techniques. Finally, an additional goal of this research is it to develop a long short-term (LSTM)-based interrogative utterance detector—a tool that can enhance language measure statistics by enabling clinicians to track a child’s response to questions from parents.