Bi Lige, Rainer Gruhn
Phoneme recognition of non-native speech
Abstract:This report examines phoneme recognition of non-native speakers. We perform
phoneme recognition with HTK HVite. A phoneme bigram provides some phonotactic
constraint. The recognition results are compared to a canonical phoneme transcription
using the DP-alignment algorithm, as implemented in HTK HResults, to get the
phoneme recognition accuracy and confusion matrixes. The confusion matrixes are
turned into graphics to visualize confusion patterns. From the analysis of confusion
matrixes and graphics, we can find which phonemes are frequently mispronounced by
speakers from different nations. The influence of the type of acoustic model is examined
by recognizing the same speech using monophone, biphone and triphone models.
Monophone models achieved highest phoneme accuracy.