Inverse Filter Parameters for Speaker Identification.
Final technical rept. 22 Apr 74-21 Apr 75,
SPEECH COMMUNICATIONS RESEARCH LAB INC SANTA BARBARA CALIF
Pagination or Media Count:
The capability to perform text-independent speaker identification is desirable because in many practical situations the text of a speech sample being tested will not be the same as the the text from which information was initially gathered about the speaker. The method of comparing feature vectors from individual sound units, usually vowels, has been studied as one possible approach to the problem. This report deals with an evaluation of three different types of parameters which are obtainable from the inverse filter method of linear prediction, namely the inverse filter coefficients, reflection coefficients, and cepstral coefficients. Using 821 samples of the vowel epsilon as in the word bed from 16 speakers a series of speaker identification experiment was performed using a Euclidean distance measure and a weighted Euclidean distance measure. Results showed that independent of the kind of experiment, or the distance measure, the cepstral coefficients and reflection coefficients had similar error rates and the errors were consistently lower than with inverse filter coefficients. It also demonstrated that these low-dimensional coefficient vectors can classify speakers nearly as well as spectral vectors which have a larger number of vector elements. The cepstral coefficients are of particular interest because they are linearly related to the log spectral domain and can be computed from the inverse filter coefficients through a recursive procedure which is computationally more efficient than the classical double FFT procedure.
- Statistics and Probability
- Voice Communications