Formant Trajectory Estimation from a Linear Least-Squares Inverse Filter Formulation

摘要

The purpose of the document is to develop in some detail a new method for accurately estimating formant trajectories from voiced speech. Given a segment or frame of speech data, an inverse filter is designed such that the input signal is transformed into a best estimate of white noise. By constraining the filter length, an estimate of the smooth resonance structure of the input signal is obtained from the resulting inverse filter spectrum. The peaks of this spectrum define the raw data for the frame being analyzed. Based upon all of the raw data for the utterance it is generally a trivial task to define the first three and sometimes four formant trajectories by inspection. Efficient computational techniques, and considerations in the application of the inverse filter are discussed in detail. Fortran programs for implementing the inverse filter are presented as appendices. (Author)