This paper deals with the problem of reducing the computational complexity of ASR algorithms for embedded systems. Particularly, three methods for simplifying the computation of state observation likelihoods of continuous density based HMMs are proposed. Feature component masking, variable-rate partial likelihood update and density pruning all result in significant savings in the decoding complexity with marginal impact on the recognition performance. A combination of feature component masking and density pruning was evaluated in a small vocabulary, 25-lingual, speaker independent, isolated word recognition system. With a computational complexity reduction of 62% compared to the baseline system, a marginal, 1.6/6.5% relative error rate increase was obtained without/with online MAP adaptation on the average in clean and noisy operating environments. The presented framework can also be extended to larger vocabulary systems.