The accuracy of speech recognition systems degrades severely when the systems are operated in adverse or noisy environments. This paper studies and presents practical results on the application of a technique for robust speech endpoint detection in the presence of additive noise. The technique uses wavelet analysis as an instrument for subband decomposition in order to compute a metric that defines the criterion for endpoint detection. It is demonstrated that this metric is robust to different types of noise such as Gaussian and car noise. Experimental results are focussing on the exploration of the ability of the proposed algorithm to give correct speech boundaries decisions as a function of the type of wavelet decomposition as well as the value of signal to noise ratio. Comparing with classical endpoint detectors this approach overcomes by far all the drawbacks and it may be considered an appropriate candidate for the application in speech recognisers working in noisy environments.
展开▼
