Using noise reduction in mode selection and pitch search

We present a novel method of exploiting noise reduction prior to mode selection in classification-based speech coding. Certain parameter estimation and mode selection is performed on a denoised signal to ideally remove the non-speech components before signal classification, although the original input is coded. This method is employed in a recent ITU-T standard, G.718, which provides state-of-the-art performance for narrowband and wideband speech between 8 and 32 kbps. We also present the pitch tracking algorithm of G.718. Open-loop pitch lags are estimated from the denoised signal.