Using SIMD technology to speed up likelihood computation in HMM-based speech recognition systems

Most state-of-the-art LVCSR systems are based on continuous density HMMs, which are typically implemented using Gaussian mixture distributions. Such statistical modeling systems usually operate slower than real-time, largely because of the heavy computational overhead of the likelihood computation. The objective of our research is to investigate application of modern SIMD technology to speed up the likelihood computation without degrading the recognition accuracy. In this paper, the likelihood computation of continuous density HMMs is analyzed to show that the conventional way of sequential computing is time-consuming and the likelihood computation itself can be implemented in parallel. A SIMD-based algorithm which can carry out parallel likelihood computation is presented in this paper. Likelihood computation modules in HTK3.4 toolkit have been modified with SIMD instructions to implement this algorithm. Experiments on TIMIT and WSJ0 corpora show that the SIMD-based data-level parallelism can significantly reduce the time overhead for likelihood computation.