A comparative analysis of the topological structures of different LPC feature-based speech models

Describes initial experimentations done on three LPC (linear predictive coding) derived feature-based speech models: the LPC-cepstrum, the LSP (line spectral pair) and the postfilter-cepstrum (PFL). A comparative analysis of the topological structures of these models is also given. The structures are basically self-organizing feature maps which accept these models as inputs and after training, used to distinguish between isolated word utterances and speakers. A small database of 5 utterances and 4 speakers is initially used. The performance index of isolated word recognition and speaker identification for all models are calculated based on a hit-and-miss ratio and are also discussed. Experimental results reveal that the three parameters are comparable in performance. The LSP has a slight edge over the other two feature vectors in distinguishing isolated words