A method using explicit prior knowledge about phonemes is presented for improving automatic speech recognition (ASR) performance in low resource settings. First, a fixed-length encoding is defined for each phoneme, where each element of the encoding represents a distinct phonetic feature. Second, a phonetic feature prediction layer is put in a deep neural network (DNN) and the feature predictions are used to make the final token predictions. Experiments are conducted in multilingual settings where Ainu is the target low-resource language. Effectiveness and robustness of this method is explored with varying amounts of training data.