HMM modeling a speech segment: The output distributions are modified considering noise and energy component coming from preceding segments due to reverberation.
The performance of a speech recognition system degrades in noisy and reverberant environment as trained HMMs no longer effectively represent speech under these conditions. We proposed three different methods to adapt HMMs by considering the effects of noise and reverberation. Besides, we proposed a feature-based method exploiting spectral dynamics in speech to reduce the effect of noise on speech features.
Dealing with noise and reverberation is unavoidable, and is a challenging problem specially when noises or the environmental characteristics change. Still there is a long way to go for a practically realizable speech recognition system that would work satisfactorily in any given environment.