HMM-based echo and announcement modeling approaches for noise suppression avoiding the problem of false triggers

In the past, we proposed an HMM-based residual echo modeling technique that proved to be e ective in eliminating false triggering and enhanced the recognition of valid keyword speech. The narrow training strategy based on single recorded prompt made the echo model a better discriminator, though it has the possible drawback of requiring retraining if the prompt changes. To overcome this intrinsic problem, we consider building a general echo model that is trained over many sentences from varied voices. The experimental results show that the multiple echo models in conjunction with suitable ller model to represent the extraneous speech not only provide good recognition accuracy but also yield better out-of-vocabulary rejection.