An ASIC design for linear predictive coding of speech signals

In the real-time speech recognition, the predictor coefficients of speech signals are used as the recognizing features and should be computed faster than the sampling rate. Under such performance constraint, the objective is to design this circuit as cheap as possible. Autocorrelation method is adopted for computing the coefficients because of its stability of the results and regular computations. A two-step pipelined functional unit is designed for calculating those regular computations. For the divisions needed in autocorrelation method, the prune-and-search approach is used to compute them. Because this approach can also be performed by the identical pipelined functional unit, the extra operator for the divisions is not required. Although the authors have designed this chip with least storage element and functional unit, however, the performance is still much faster than the real-time request.<
>