Non-steady state speech analysis method with dynamic feature enhancing effect

Extraction of speaker independent features, which make the separation of /b/ /d/ /g/ or /p/ /t/ /k/ possible, is one of the most difficult problem in automatic speech recognition. The authors propose a new speech analysis method to handle these sounds characterized by high speed articulations. The method is based on an autoregressive model with linearly time variant parameters in the analysis window. Recursive method, which is achieved by solving simultaneous linear equations with same number of parameters as in LPC, is proposed assuming the framewise continuation of each parameter. An articulatory dynamic feature enhancing effect is created by the introduction of vocal tract reflection coefficients and the enhancement of the vocal tract (acoustic tube) shape change between adjacent frames. In experiments on Japanese CV-syllables, where C expresses stops, comparisons have been made between the proposed method and LPC-based method, and widely different results were obtained for the transient parts, especially in labio-dental sounds such as /d/.