Speech conversion from clean conditions to telephone ones

Abstract

This paper addresses the application in speech recognition of simulated telephone speech, which is generated from clean speech by approximately mimicking actual telephone channel conditions. Maximum Likelihood Linear Regression (MLLR) algorithm was performed to conduct experiments on evaluating the performances of HMM recognizers, which were trained from clean speech and from generated telephone data, respectively. The test and adaptation data were recorded by piping clean speech through local telephone network. The experiments without adaptation report that the simulation models trained on generated data can give an obviously higher rate than the clean speech. The adaptation performances show that MLLR lends itself to further improve the recognition performance of telephone recognition system. The results show that telephone speech recognition performance can be effectively improved using the generated data, and its generating method can reduce the acoustic mismatch between training and testing data that was induced by the shortage of actual telephone speech.