On the Application of Energy Contours to the Recognition of Connected Word Sequences
L. R. Rabiner · AT&T Bell Laboratories Technical Journal · 1984
It has recently been shown that small but consistent improvements in isolated word recognition accuracy can be obtained by supplementing the Linear Predictive Coding (LPC) features for each frame of a word by a normalized energy value for that frame. The key idea in using energy is to normalize the frame energy by the local energy maximum in time (i.e., relative to the peak energy of the spoken word). If we want to extend the concept of using frame energy as a supplement to the LPC feature set for connected word recognition, we must provide a dynamic method of energy normalization so that the peak energy within strings can closely approximate the energy contours of individual words strung together. In this paper such a dynamic energy normalization is proposed, and it is shown to provide improvements in connected word recognition applications. The normalization consists of determining a continuous peak energy contour for the speech, where the peak energy is determined over periods of time essentially corresponding to a syllable, and then modifying the actual energy contour with the peak energy contour so that absolute energy maxima occur about once per syllable. In this manner, the dynamically normalized, temporal energy contour of the word string effectively provides a set of temporal markers of high-energy events (content words) that aid the recognition of connected word sequences.