Comparison Of Speech Detection By Many Pre-processing Methods

The domain of speech signal processing is expansive. Speech signal processing has a broad prospect in the application in reality. Speech detection is an aspect of speech signal processing and is important to speech recognition. The noise is single or in accordance with some hypothesis in current speech detection. To avoid singleness of the noise, the sort of nonspeech is many in this dissertation for example the sound of water, car, animals and laugh and so on.Linear prediction code and Mel-scale cepstrum coefficients are effective method in speech feature extraction and important in speech signal processing. A new feature extraction method is presented in this dissertation and is proved practicability. The main jobs in this dissertation include follows:1. The theory of LPC and Mel-scaled cepstrum analysis is introduced in this dissertation and how to extract LPCC and MFCC is elaborated.2. Some theory of high-dimension space is introduced and a new feature extraction method avito correlation-angle is presented based on high-dimension space geometry.3. Based on high-dimension space geometry, every speech sample is looked as a point in space. Then the speech sample point is extracted feature by LPC, Mel-scaled cepstrum analysis or auto correlation-angle. Their feature is looked as a point too. The feature point projects on coordinates then discriminated speech between nonspeech by the projection result. The merits and weakness of auto correlation-angle are compared with LPC and Mel-scaled cepstrum analysis.