Two important classification problems in speech signal processing are speech detection and recognition. They are easily affected by noisy environments where there may exist concurrent noises due to movements of desks, door slams, etc. To solve this problem, a linear-discriminant-analysis-based type-2 fuzzy-neural-network (LDA2FNN) is proposed. In handing noisy data with uncertainties, type-2 fuzzy-systems generally outperform their type-1 counterparts. Therefore, type-2 fuzzy-sets are used in the antecedent parts to cope with the noisy data. The most important consideration for classification problems is the "discriminability". To increase the "discriminability", linear-discriminant-analysis (LDA) is applied in the consequent parts. Compared with other existing fuzzy neural networks, the novelty of the proposed LDA2FNN is its consideration of both uncertainty and discriminability. Furthermore, its computation load is low. In experiments, LDA2FNN is successfully applied to speech detection and recognition. Experimental results indicate that the proposed LDA2FNN performs better than the other fuzzy neural networks.