Statistical approach to enhancing esophageal speech based on Gaussian mixture models