Warped-twice minimum variance distortionless response spectral estimation

摘要: This paper describes a novel extension to warped minimum variance distortionless response (MVDR) spectral estimation which allows steer the resolution of envelope lower or higher frequencies while keeping overall estimate and frequency axis fixed. effect can be achieved by introduction second bilinear transformation warped-MVDR estimation, but now in domain as opposed first is applied time domain, compensation step adjust for pre-emphasis both transformations. In feature extraction process an automatic speech recognition system this emphasize classification relevant characteristics dropping irrelevant features according signal analyze, e.g. vowels fricatives have different therefore should treated differently. We compared on evaluation data Rich Transcription 2005 Spring Meeting Recognition Evaluation got word error rate reduction from 28.2% 27.5%.