Parallel MLP Feature Extraction for Speech Recognition

Nelson Morgan, Chris Oei1 and Adam Janin2

International Computer Science Institute and Microsoft/Intel Parallel Computing Laboratory

There is much parallelism inherent in speech recognition techniques, particularly for the systems built at ICSI. The front-end for acoustic feature extraction performs spectral analysis of the audio signal and classifies phonetic units with a multi-layer perceptron (neural network); the bulk of this computation is dense matrix-matrix multiplication, for which we can exploit BLAS subroutines. Using multi-threaded BLAS libraries enabled significant speedup on MLP forward pass and backpropagation training on a multicore CPU architecture. We hope to gain even more speedup by exploiting the greater parallelism of a GPU architecture. We are also investigating feature extraction schemes in which many MLPs operate independently in parallel.