Automatic recognition of vowel length in Japanese for a CALL system motivated by perceptual experiments

Acquisition of the Japanese vowel length contrast can be problematic for non-native speakers. For these speakers, a CALL system which can automatically recognize vowel length could be of great benefit for pointing out their errors and issuing corrective feedback. However, a method that can adequately do this has not been proposed yet. Vowel length recognition is made difficult because the vowel length distinction is dependent on the surrounding vowel durations which vary due to speaking rate and other factors. Hidden Markov Models (HMMs), the standard way of recognizing this distinction, do not make use of this information. Methods have been proposed to recognize this in the past, but they do not appear viable unless knowledge about the durations of other vowels is present. Thus, we carry out perceptual experiments to gain more knowledge about the vowel length contrast. From this analysis, we develop an automatic recognition algorithm for vowel length that uses support vector machines (SVMs). We tested this method on a speaking rate corpus, native speech, and non-native speech. The method produced recognition results that are overall superior to HMMs and also more robust against speaking rate differences with an average of a 0.83 correct recognition rate for the 3 datasets. The error and non-error classification rates on non-native speech for this are 0.86 and 0.84 respectively.