Efficient nn-rasud search space reduction in a large vocabulary speech recognition system

In very large vocabulary speech recognition systems using the hypothesis-veriiicali~)np aradigm, the vcrification stage is usually the most time consumiilg. State ofihi. art systems combine fixed six hypothesized search spaces with odvanccd pmninp techniques. In this paper we propose R novel strategy to dynamically calculate thc hypotliesired search spacc, rising neural neiamorks as the estimation module and designing the inpui feature set wiih a cnrcfiil gredy-based selectioii approach. The main achievement lias been a statistically signilicaiit rclniivc dccrrase in error rate of 33.53%, while gelling a relative decrease in average computational demands of up to 14.40%.