β Scribed by Qifang Zhao; Tetsuya Shimamura; Jouji Suzuki
No coin nor oath required. For personal study only.
When the prediction coefficient is determined from the speech signal containing a white noise, using the autocorrelation method, the effect of the noise concentrates in the neighborhood of k = 0 of the autocorrelation function Rk. Theoretically, the effect of the noise is compensated and the noise immunity of the LPC analysis is improved by subtracting the noise power from the autocorrelation function R(0). When conventional noise compensation is applied, however, it may happen that the LPC filter becomes unstable due to overcompensation of the noise power. This makes practical application difficult. Shimamura and coworkers, on the other hand, proposed an improved noisecompensated AR coefficient estimation method based on an iterative procedure, in order to solve a similar problem in spectrum estimation. This paper applies their improved noise-compensated AR coefficient estimation method to noise reduction of LPC speech analysis. As a result of the evaluation experiment, it is seen that this approach is effective for speech without preemphasis, but is not effective for speech with preemphasis. It is shown by a theoretical analysis that the reason for the ineffectiveness is that the effect of the noise is extended to the neighborhood of k = 1 by preemphasis. An iterative algorithm which is not affected by preemphasis is derived, where the noise power is reduced from the neighborhoods of both R(0) and R(1). The effectiveness of the method is verified by a computer simulation. The method is further extended, and an improved method is proposed that can cope with noise other than white noise.
π SIMILAR VOLUMES