Automatic speech recognition of naturalistic recordings in families with children who are hard of hearing

Abstract

Performance of an automatic speech recognition (ASR) system [LENA Research Foundation, Boulder, CO] has been reported for naturalistic, whole day recordings collected in families with typically developing (TD) children. This report examines ASR performance of the LENA system in families with children who are hard-of-hearing (HH). Machine-labeled segments were compared with human judges' assessment of talker identity (child, mother, or father), and recordings from families with TD children were compared with families with HH children. Classification models were fit to several acoustic variables to assess decision process differences between machine and human labels and between TD and HH groups. Accuracy and error of both machine and human performance is reported. Results may be useful to improve implementation and interpretation of ASR techniques in terms of special populations such as children with hearing loss. Findings also have implications for very large database applications of unsupervised ASR, especially its application to naturalistic acoustic data.