A real-time ADPCM encoder using variable order prediction

The compression of digitally encoded speech in real-time is of interest to those working on voice transmission and storage systems. This paper evaluates the application of a variable order predictor to Adaptive Differential Pulse Code Modulation (ADPCM) to lower the bit rate without mitigating the speech quality. The technique encodes the speech waveform by using a lattice prediction filter to predict the input speech waveform and then adaptively quantizing the prediction error. The prediction error is monitored and the minimum order predictor filter is determined for a prescribed error for each analysis speech segment. The order selection algorithm is derived and a program flow chart given for the fixed point implementation. By using the minimum order filter, one reduces the number of coefficients that have to be coded. A comparison is then done to contrast the new variable order algorithm with the more conventional fixed order LPC algorithm. Results include the favorable SNR performance of the coder for various bit rates in the mid-band (10-32 Kb/s) range as well as the coding aspects on the Digital Signal Processing Chip (TMS320). Several programming innovations such as adaptive bit allocation for correlation coefficient normalization are also described. Work is now in program to extend this research to lower bit rates and includes the application of noiseless compression techniques on the encoded residual.