Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


Comparing Various Feature Vectors in Automatic Speech Recognition

Paul van Alphen (1,2), Louis C. W. Pols (2)

(1) PTT Research Laboratories, Leidschendam, The Netherlands
(2) Institute of Phonetic Sciences, University of Amsterdam, Amsterdam, The Netherlands

In speech recognition suitable feature extraction is an important issue. In our research we tested the performance of several individual feature vectors, as well as various combinations of these vectors. As a basis for the feature vectors we use two types of analysis: a 15-band filterbank and an LPC-cepstral analysis. From these base-feature vectors several other feature vectors are derived. In the recognition part of our system (REXY) we used HMM's to model phone-like-units (PLU's). The system is trained on continuous speech of one male speaker. We tested the different feature vectors separately, and some of their combinations (combining feature vectors is done at a probabilistic level). The overall performance of the filterbank analysis turns out to be better than the LPC-cepstral analysis. keywords; speech recognition, HMM, feature vectors.

Full Paper

Bibliographic reference.  Alphen, Paul van / Pols, Louis C. W. (1991): "Comparing various feature vectors in automatic speech recognition", In EUROSPEECH-1991, 533-536.