Handling convolutional noise in missing data automatic speech recognition

摘要

Missing Data Techniques have already shown their effectivenessin dealing with additive noise in automatic speech recognition systems.For real-life deployments, a compensation for linear filteringdistortions is also required. Channel compensation in speechrecognition typically involves estimating an additive shift in thelog-spectral or cepstral domain. This paper explores a maximumlikelihood technique to estimate this model offset while some dataare missing. Recognition experiments on the Aurora2 recognitiontask demonstrate the effectiveness of this technique. In particular,we show that our method is more accurate than previouslypublished methods and can handle narrow-band data.