"Noise cancellation studies using a least-squares lattice filter"

The operation of a least-squares lattice filter, configured for noise cancellation, is described. The algorithm employed uses both apriori and a-posteriori unnormalised residuals and requires no auxiliary scalars. A 100-stage version of this filter has been applied to data of 5 and 10 seconds in length, sampled at 8kHz, comprising a passage of speech corrupted by additive noise. The algorithm successfully effected a noise reduction close to the optimum attainable level. By exploiting redundancy in the number of linear prediction lattice stages required, a reduction from 24 operations per stage to 9 operations was possible for the majority of stages, with no appreciable loss of performance.