Presents novel sequential and parallel learning techniques for codebook design in vector quantizers using neural network approaches. These techniques are used in the training phase of the vector quantizer design. These learning techniques combine the split-and-cluster methodology of the traditional vector quantizer design with neural learning, and lead to better quantizer design (with fewer distortions). The sequential learning approach overcomes the code word underutilization problem of the competitive learning network. As a result, this network only requires partial or zero updating, as opposed to full neighbor updating as needed in the self organizing feature map. The parallel learning network, while satisfying the above characteristics, also leads to parallel learning of the codewords. The parallel learning technique can be used for faster codebook design in a multiprocessor environment. It is shown that this sequential learning scheme can sometimes outperform the traditional LBG algorithm, while the parallel learning scheme performs very close to the LGB and the sequential learning algorithms.>

In this paper we present a study on the performance of a speaker-dependent continuous speech recognition algorithm in various background noise levels, including the mismatch tolerance of the algorithm. This mismatch exists in most applications where the user is trained in one noise level and does the recognition in different and highly variable noise levels. Finally, we introduce a pre-processing technique called 'twicing' and a simple 3-point moving average post-processor. The twicing algorithm, while still maintaining the high performance in the quiet background, reduces the error rate in the noisy environments by a significant amount. The results indicate that smoothing in addition to twicing provides significant performance improvement in high noise background.