High quality low delay wideband speech coding at 16 kb/s

Publisher Summary The technique of backward LPC prediction was investigated over wideband speech. It was found that for relatively short excitation vector lengths, the technique can produce high-quality speech with little or no distortion. Unfortunately, this typically resulted in a high overall bit rate that was primarily because of frequent updating of the excitation parameter set. To reduce the overall bit rate, it is necessary that the length of the excitation vector is increased. However, this resulted in less accurate modeling of the reference signal and as a consequence, the backward LPC prediction was performed over a noisy synthetic signal. This caused the predicted spectrum to be distorted, especially over the higher frequencies. This was primarily caused by the noise level immersing these lower magnitude frequency components—that is, the high frequencies. As a result of this spectral distortion, the processed speech became distorted resulting in a quality that was unacceptable for a high-quality wideband system. This distortion was controlled by splitting the wideband speech into two narrowband signals. The lower subband signal was still coded using backward LPC prediction.