Adaptive filtering for speech enhancement
P. M. Clarkson, P.R. White, J.A. Mardell · The Journal of the Acoustical Society of America · 1986
Several authors have considered the use of adaptive noise cancellation for speech enhancement. These efforts have concentrated on the simple noisy gradient descent (LMS) algorithm and, in the absence of an external reference measurement, have formed a reference internally using a delayed version of the input. Limited improvements in SNR (for voiced speech enhancement) have resulted due to the inability of the filter to cope effectively with variations in power (of the speech) across frequency, discrete changes in the pitch period and because of the limited SNR of the reference. This study has approached the problem by replacing the noisy reference measurement by a synthetic reference which consists of a low-pass filtered series of discrete impulses at intervals corresponding to the pitch period of the input. This method has been compared with a modified version of the usual noise canceller (using an exact least-squares algorithm) on a sustained vowel segment. Higher SNR improvements for the synthetic reference have been measured. However, the method does require an estimate of the pitch period and appears quite sensitive to errors in this estimate.