In this paper, we propose a perceptually-motivated method formodifying the speech power spectrum to obtain a set of linearprediction coding (LPC) parameters that possess good noiserobustnessproperties in network speech recognition. Speechrecognition experiments were performed to compare the accuracyobtained from MFCC features extracted from AMR-codedspeech that use these modified LPC parameters, as well as fromLPCCs extracted from AMR bitstream parameters. The resultsshow that when using the proposed LP analysis method, therecognition performance was on average 1.2% - 6.1% betterthan when using the conventional LP method, depending on therecognition task.
展开▼