A left-to-right algorithm for analyzing pitch contours and estimating their underlying representation as a sequence of parametered commands is proposed. It is based on a generalization of recursive least squares optimization to a nonlinear model. This parameter fitting is embedded in a search guided by prosodic event detection, which determines when a candidate structure has to be augmented to keep fitting incoming data. The algorithm applied to Japanese read sentences successfully estimated 91% of lexical pitch accent positions.
Keywords: prosody, intonation, speech recognition.