Statistical Block Protection Coding for DPCM-Encoded Speech

Block protection coding, whereby a block of data words is protected by the addition of special code words or letters, is a common feature in communication systems for noisy channels. In algebraic error detection and correction, for example, the protection is derived from parity check bits. The number of parity checks, and hence the redundancy, increases with the number of data bits protected, but the resulting error-coding procedures are quite general, being applicable to any type of data, irrespective of its source. Nevertheless, with sources such as speech, where it is not crucial to recover every speech-carrying bit without error, it is meaningful to look for certain special, compact forms of non-algebraic block protection. The idea is to transmit a protection word that identifies some perceptually significant parameter of a speech-waveform segment; knowing the (correct) value of this parameter, the receiver can perform error-detecting and error-correcting operations, which may be only partial in an algebraic sense (due to the compactness of the protecting procedure) but nevertheless quite adequate from a speech-perception viewpoint. In one recent investigation1 along these lines, each block of differential PCM words was protected by a reference PCM word that signified the speech amplitude at the end of the block. Error detection was 1647