Fixed-point truncation arithmetic implementation of a linear prediction autocorrelation vocoder

The implementation of the Linear Prediction Auto-correlation Vocoder (LPAV) with fixed-point truncation arithmetic is presented. For a speech bandwidth of up to 3.3 kHz, it is possible to implement LPAV with a maximum of 16-bit fixed-point arithmetic, thus allowing for very efficient software simulation on minicomputers or economical hardware implementation on a special purpose computer. The resultant degradation with respect to a floating-point implementation is negligible. The three most critical system components are discussed in detail along with various tradeoffs such as effects on system stability as a function of preemphasis and sampling frequency. The results are also applicable to general finite word length digital filter implementation problems.