Implementation of recovery of speech with missing samples on a DSP chip

Abstract

FFT block codes have been proposed as an error correction code similar to RS codes. In the Letter this novel technique is applied to speech signals with erasures. A convolutional method is also implemented which is capable of reconstructing oversampled speech signals with a certain percentage of lost samples, given that the average sampling rate remains higher than the Nyquist rate. The convolutional method uses FIR digital filters in the time domain as opposed to lowpass filtering the FFT block in the block code. The system is implemented in real time on the floating point DSP processor AT&T DSP32C. The quality of speech for the various techniques is evaluated both subjectively and objectively.