Speaker adaptation of stochastic segment models using Maximum Likelihood Linear Regression

The stochastic segment model (SSM) has been shown to be a competitive alternative to the hidden Markov model (HMM). In this paper, we extend the theory of Maximum Likelihood Linear Regression adaptation (MLLR), which is widely used in HMM-based system, to the stochastic segment model, and derive the SSM-based MLLR adaptation method. Continuous speech recognition experiment using the SSM-based MLLR adaptation derives about 7.5% relative improvement from the speaker independent (SI) system and shows the SSM-based MLLR adaptation can also improve the recognition performance just as the HMM-based MLLR adaptation does.