Multi-band approach has recently been introduced for recognition of speech corrupted by frequency-localized noise, showing higher robustness than the traditional full-band approach. However, the multiband approach has been found to be less robust for wide-band noise than the full-band approach. In this paper, we present a multi-band recognition system based on the combination of the probabilistic union model and the frequency-filtering technique. The probabilistic union model is used to combine the features from the individual sub-bands without requiring information about the sub-band corruption. The frequency-filtering technique is used to produce the feature vector for each sub-band, which is similar to the usual cepstral feature but does not spead the frequency-localized noise over the sub-bands. We demonstrate that this combination results in a system that is equally effective for dealing with both frequency-localized noise and wide-band noise.