Reconsidering the method of finding phonemic category boundary in speech production

Abstract

Several previous studies in speech production obtained a boundary of two phoneme categories by an exhaustive search for the "optimal boundary" that provides the best phoneme classification on an acoustic dimension such as voice onset time. However, they used this method without confirming its accuracy in estimating the true boundary. The present study examined this accuracy by conducting a Monte Carlo simulation under several conditions of the means and variances of two phoneme categories. The results revealed that, compared with the statistically estimated boundary, the "optimal boundary" tended to have bigger root mean square errors from the true category boundary. This finding indicates that the "optimal boundary" is less accurate than the statistically estimated boundary for estimating the true boundary. Our study considers the limitations of the previous method and recommends a statistically estimated boundary in future investigations.