ESSV Konferenz Elektronische Sprachsignalverarbeitung

Title: The impact of correlated features in speech recognition

Authors: Harald Höge

Abstract:

Hidden Markov Models assume that adjacent feature vectors are statistically independent. Yet the use Δ and ΔΔ operations, super-vectors and LDA leads to highly correlated feature vectors contradicting the independence assumptions. Experiments [1] have shown, that more sophistical acoustical models lead to no substantial decrease in error rate. In order to investigate these findings we use simulated feature vectors having probability distributions similar as derived from real speech data. The used distribution model exactly the statistical properties of adjacent feature vectors. We made recognition experiments on 607 segments derived from tri-phones, which were realized by two feature vectors. The experiments confirm that the use of second order statistics does not improve the recognition rate substantially. Further we can show with simulated features that the error rate decrease with decreasing degree of correlation.


Year: 2011
In session: Sprachsignalverarbeitung, Spracherkennung und Sprachsynthese I
Pages: 31 to 38