Prosody is the music of speech, the high and low tones accompanying each word that give nuance to what we say. Low pitch accents (L*) suggest that the accompanying word is shared information, while high tones (H*) suggest that the word conveys new information. The pitch information that carries these tones is usually thrown out in speech recognition, but it seems as though they might be pretty useful for a computer to understand!

In joint work with Eli Kim (Yale), we developed a simple system for finding high and low tones in speech, and used it to analyze recordings from the CHILDES corpus of child-directed speech. We found that simple high and low accents did tend to appear in roughly the places we expected (paper), suggesting that this could be an extra signal that children use to develop the concept of beliefs different from theirs.

Working with Orit Shaer of Wellesley, we hope to develop an automated system for detecting H* accents in group brainstorming sessions, so that the key words could be pulled out for easy retrieval later. A system that automatically finds bursts of H* accents could be useful in later bookmarking important moments in these sessions. We hope to integrate this tool into a comprehensive system for helping to guide, analyze, and index brainstorming sessions. (Image courtesy of Wellesley HCI Lab.)