Least Significant Bit Coding of Speech

We investigate the application of distributed source coding to a single-channel speech or audio signal. We compare two methods. In the first, we encode signal sequences using low-density parity check codes in an approach that resembles methods used in video coding. In the second, we encode each sample with only a set of least significant bits (LSB) of the quantization index. In both methods the parameters of a statistical signal model are estimated at the encoder and transmitted to the decoder. The signal model is used only at the decoder. Our results indicate that the LSB coding method is more effective. We show that the order-k entropy of a sequence of LSBs converges more quickly to the entropy rate of the scalar quantized signal than the order-k entropy of the quantized signal itself. As a result, we can approach the rate-distortion bound by applying entropy coding to successive short (several samples) segments of the sequence of LSB sample. The LSB coding method can be extended to lattice quantization.