The influx of noise through inlet streams is often a problem in the operation of large-scale microbial fermentations. It can distort the otherwise smooth performance and, more seriously, displace the fermentation to an undesirable state. Therefore, removal or reduction of the noise content of measured data is important for retrieving the true process variables for bioreactor operation and control. This is done by noise filters, which are soft devices that process noisy data and generate less noisy values with identifiable features. Three types of filters have been compared here by applying them to a continuous fermentation by Saccharomyces cerevisiae under (a) monotonic, (b) oscillating and (c) chaotic operation. Recognising self-similarity as a characteristic feature under the influence of noise, fractal dimensions of the output concentrations are suggested as effective indexes of both noise-affected and noise-filtered performance. On this basis, a hybrid neural filter (HNF) was the best, an auto-associative neural filter (ANF) was somewhat inferior and an extended Kalman filter (EKF) the poorest. While these results and similar observations for other microbial systems favour the use of both fractal dimensions and the HNF, the EKF and other algorithmic filters have some merits, which are discussed.