Harnessing graphics processors for the fast computation of acoustic likelihoods in speech recognition