Optimization of voiced/Unvoiced decisions in nonstationary noise environments

This paper describes a way of optimizing the autocorrelation method of voiced/unvoiced decision for speech which is heavily degraded by nonstationary ambient noise. Usually a constant threshold is adopted to which the correlation peak value is compared for voiced/ unvoiced decision. The optimal threshold is, however, a function of noise characteristics and the signal-to-noise ratio. This paper presents a method of estimating the probability density function of correlation peak values from noisy speech and also of estimating the optimal threshold based on the expected error rate of the voiced/unvoiced decision. The performance of the proposed method has been tested under various noise characteristics and signal-to-noise ratios. These tests show that the estimated threshold is very close to the true optimal threshold in almost all cases. The proposed method also retains the optimality under slowly time-varying noise conditions, even if no a priori information is given about noise characteristics or noise level.