A cross-language vowel normalisation procedure

Abstract

A variation of log-mean normalization that can be applied in cross-language and cross-dialect experiments, is presented. Vowel normalization procedures seek to remove inter-speaker variance due to factors such as vocal tract size, which human listeners discount while identifying vowels. It is expected that the mean vocal-tract length are approximately equal over large and sufficiently balanced samples of speakers. The cross-language normalization procedure is tested using acoustic data from productions of L1-Spanish non-low front vowels and L1 English non-low front vowels. Three models are trained and tested using non-normalized log-Hertz values, language-normalized values, and cross-language normalized values respectively. The models trained on normalized L1-English vowels had a slightly higher correct-classification rate on the training data than the model trained on non-normalized data. The vowel normalization increased the correlation between monolingual English listeners.