Variable Rate ADPCM Based on Explicit Noise Coding

Multiple-stage coding, where the reconstruction noise from an initial stage is itself coded for transmission in a subsequent stage, is known to provide substantial gains over single-stage coding in the context of deltamodulation using oversampled inputs.1,2 In this paper, we consider two-stage systems for multibit differential pulse code modulation with adaptive quantization (ADPCM) coding of Nyquist-sampled speech inputs. Unlike systems that permit oversampling, signal-to-noise ratio 657 (s/n) gains in our systems will be seen to be either slightly negative, or positive but nondramatic. However, the proposed systems have a feature that is common to all noise-coding systems, the property of embedded coding: the output bit sequence of the coder contains a subsequence that can be used in a straightforward manner to provide lower bit-rate operation with an output speech quality very close to that of conventional operation at the lower bit rate; as a result, the channel or receiver can switch, as needed, between low-rate and highrate modes. The possibility of variable-rate operation is a very desirable feature in digital communication systems such as packet-switched voice networks. 3 A PCM coder is inherently an embedded coder. Least significant bits in a PCM codeword can be progressively dropped, with a graceful loss of quality that is no greater than about 6 dB/bit. Conventional differential pulse code modulation (DPCM) is not an embedded coding system in a similar sense because of the presence of a feedback loop in coder and decoder.