Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

On the Use of Features from Prediction Residual Signals in Speaker Identification

Jialong He, Li Liu, GŁnther Palm

Department of neural information processing, University of Ulm, Germany

A by-product of the LPC analysis is the generation of a prediction residual signal e(n). e(n) is usually ignored in the major applications of speech analysis, and only the LPC coefficients or parameters derived from the LPC coefficients are used to compose feature vectors. Since e(n) carries all information that has not been captured by the LPC coefficients, an algorithm was proposed to calculate parameters from this prediction residual signal. The effectiveness of these features (named as RCEP coefficients) for speaker identification was evaluated. This approach yielded promising results. In an evaluation experiment in which the learning vector quantization (LVQ) networks served as classifiers, the correct identification rate for 112 male speakers was 88.8% for feature vectors composed of LPC based cepstrum (LPCC) alone, but reached 96.9% when the LPCC coefficients were combined with the RCEP coefficients.

Full Paper

Bibliographic reference.  He, Jialong / Liu, Li / Palm, GŁnther (1995): "On the use of features from prediction residual signals in speaker identification", In EUROSPEECH-1995, 313-316.