Statistical modeling of infant-directed versus adult-directed speech: Insights from speech recognition
Katrin Kirchhoff, Steven M. Schimmel · The Journal of the Acoustical Society of America · 2003
Studies on infant speech perception have shown that infant-directed speech (motherese) exhibits exaggerated acoustic properties, which are assumed to guide infants in the acquisition of phonemic categories. Training an automatic speech recognizer on such data might similarly lead to improved performance since classes can be expected to be more clearly separated in the training material. This claim was tested by training automatic speech recognizers on adult-directed (AD) versus infant-directed (ID) speech and testing them under identical versus mismatched conditions. 32 mother–infant conversations and 32 mother–adult conversations were used as training and test data. Both sets of conversations included a set of cue words containing unreduced vowels (e.g., sheep, boot, top, etc.), which mothers were encouraged to use repeatedly. Experiments on continuous speech recognition of the entire data set showed that recognizers trained on infant-directed speech did perform significantly better than those trained on adult-directed speech. However, isolated word recognition experiments focusing on the above-mentioned cue words showed that the drop in performance of the ID-trained speech recognizer on AD test speech was significantly smaller than vice versa, suggesting that speech with over-emphasized phonetic contrasts may indeed constitute better training material for speech recognition. [Work supported by CMBL, University of Washington.]