Signal adaptive spectral envelope estimation for robust speech recognition

摘要: This paper describes a novel spectral envelope estimation technique which adapts to the characteristics of observed signal. is possible via introduction second bilinear transformation into warped minimum variance distortionless response (MVDR) estimation. As opposed first transformation, however, applied in time domain, must be frequency domain. extension enables resolution estimate steered lower or higher frequencies, while keeping overall and axis fixed. When embedded feature extraction process an automatic speech recognition system, it provides for emphasis features that are relevant robust classification, simultaneously suppressing irrelevant classification. The change may steered, each observation window, by normalized autocorrelation coefficient. To evaluate proposed adaptive technique, dubbed warped-twice MVDR, we use two objective functions: class separability word error rate. Our test set consists development evaluation data as provided NIST Rich Transcription 2005 Spring Meeting Recognition Evaluation. For both measures, consistent improvements several speaker-to-microphone distances. In average, over all distances, front-end reduces rate 4% relative compared widely used mel-frequency cepstral coefficients well perceptual linear prediction.