A non-iterative algorithm for obtaining multi-pulse excitation for linear-predictive speech coders

Many commercial applications of vocoders require good-quality speech reproduction at moderate bit-rates and at low cost. Multi-pulse-excited linear prediction provides excellent quality of speech reproduction at moderate bit-rates. An analysis-by-synthesis method for deriving the multi-pulse excitation sequence is described in [1]. It employs minimization of the spectrally-weighted energy in the synthesis error signal. The complexity of this method is judged to render it unsuitable for low-cost implementation at present. A modified algorithm is derived, exploiting the time-domain behaviour of the spectral weighting function. The modified algorithm has been used to implement a 7.2 kbits/s real-time vocoder on a double Eurosize circuit board (160mm × 233.4mm).