Abstract: A neural network architecture, K-subspaces and time-delay auto-associators, is proposed for phoneme recognition. It extends the phoneme filter neural networks approach by adding linear auto-associators to create p-dimension subspace, and an iteration is employed to improve the decision. It is good to capture the time-sequence information in speech signal. The architecture proposed could provide a high recognition performance without traditional neural network's shortcoming. Some recognition simulations for both English and Chinese phonemes are conducted, and the recognition rate is 84.38% which is better than phoneme filter neural networks approach.