Noise carried by the feed stream is a common feature of large-scale bioreactor operations. This noise may be modelled by a set of time-dependent Gaussian distributions. Recent studies have shown that neither unfiltered nor completely filtered noise is desirable. The best performance requires optimally filtered noise. A previous publication in this journal showed that streptokinase (SK) activity in a fed-batch fermentation can be improved substantially through controlled static filtering. Later work with beta -galactosidase showed that dynamic filtering by means of a neural network was superior, especially when it was coupled to a neural filter. That concept has been applied to SK. Coupling of two neural networks increased the peak SK activity (in g/l) by 42% over that for a noise-free feed whereas the improvement with a static filter was 22%.