An investigation of spectral match statistics using a phonemically marked data base

With the current popularity of template matching for speech recognition systems, it is important to have the strongest possible spectral match statistic for scoring template frames against speech frames. This paper describes a phoneme based technique that has proven useful in studying match statistics and compares a new filter bank based statistic derived with this technique to several linear prediction and filter bank statistics. Two filter bank based statistics, the new statistic and a symmetric variation of the distance measure proposed by Klatt [5] performed better than the linear prediction statistics in a word recognition task on continuous speech.