Automated episode selection of child continuous speech via blind source extraction
Marisha L. Speights, Keith D. Gilbert, Joel M. MacAuslan, Richard S. Goldhor · The Journal of the Acoustical Society of America · 2018
Commonly, in both research and clinical conditions in which speech recordings are collected from children, the non-targeted adult speech and the targeted child speech are simultaneously recorded during conversational turn-taking tasks, and the utterances of all talkers are audible on every channel of the recording. As a result, the intended analysis of the targeted speaker becomes encumbered by the requirement for a human analyst to identify and mark the episodes in the recording where the targeted speaker is the only active sound source. Blind source extraction (BSE) techniques are utilized here to automatically deliver the isolated child’s speech, even when the child’s and adult’s speech overlap, or when extraneous interfering noise is present. To test the performance of the BSE methods for automated episode selection in child continuous speech, ten typically developing children were each recorded speaking thirty-three sentences in the turn-taking scenario with four microphones placed in the environment. Correlation between the numbers of human-marked episodes identifying the target speakers to the machine-selected episodes was examined. Results of the Pearson correlation indicate there is a significant positive association between hand counts and BSE machine counts (r = 0.85, p <0.01).