Inverse filter for speaker identification

The objective of a speaker identification task is to determine if a sample of speech from an unknown talker can be associated with a sample of speech from one of several reference talkers. Of the many different techniques for performing speaker identification, the method of comparing feature vectors from individual sound elements was chosen for identification experiments. The experiments utilized the inverse filter analysis technique, which is also described as the autocorrelation method of linear prediction. Text-dependent identification experiments were performed with a population of ten male talkers whose speech samples were bandlimited to 3250 Hz. Thirteen sound units, 10 vowels and 3 nasals, were chosen as the elements from which feature vectors would be extracted. Many additional identification tests were performed for the purpose of evaluating different distance functions and different feature vectors. (Modified author abstract)