Noise elimination in degraded Kannada speech signal for Speech Recognition
Thimmaraja Yadava G, Jai Prakash T S, H. S. Jayanna · 2015
In this paper, we demonstrate the methods for preprocessing of noisy speech data to build an Automatic Speech Recognition (ASR) for Kannada language. The methods are spectral subtraction with Voice Activity Detection (VAD), Linear Prediction Coefficient (LPC) analysis of speech using autocorrelation and periodogram subtraction method. In spectral subtraction method, noisy speech data is segmented and windowed into 50% overlapped frames and is processed frame by frame. An application of VAD is to detect only active regions of speech signal. In LPC analysis of noisy speech using periodogram and autocorrelation subtraction methods, the autocorrelation coefficients are calculated first and then by subtracting the periodograms of additive noisy signal from corrupted speech signal, the noise is eliminated.