Noisy Speech Signal Processing using Autocorrelation Function

T. Sakai, T. Murakami, and Y. Ishida (Japan)


SNR, linear prediction analysis, autocorrelation function, formant frequency, power spectrum, and spectrum envelope.


In this paper, we propose a new spectrum extraction method that is not affected by the noise. It is based on the fact that the autocorrelation function decreases the influence of the noise. The autocorrelation function has the following characteristics. 1) The white noise element concentrates on a zero delay part. 2) The spectrum is equivalent to the power spectrum of the audio signal. Therefore, it is possible to estimate accurately the spectrum envelope and the formant frequency by using the autocorrelation function instead of the audio signal. In the proposed method, we first estimate the noise spectrum from the unvoiced section and determine the threshold level from the peak value of the noise spectrum. The power spectrum is lifted up by this threshold to concentrate the influence of the noise on the zero delay part of the autocorrelation function. The formant frequency and the spectrum envelope are calculated by applying the linear prediction analysis to the autocorrelation function to decrease the influence of the noise.

Important Links:

Go Back