Recently, attention based sequence-to-sequence model become a research hotspot in speech recognition. The attention model has the problem of slow convergence and poor robustness. In this paper, a model that jointed a bottleneck feature extraction network and attention model is proposed. The model is composed of a Deep Belief Network as bottleneck feature extraction network and an attention-based encoder-decoder model. DBN can store the priori information from Hidden Markov Model so that increasing convergence speed of and enhancing both robustness and discrimination of features. Attention model utilizes the temporal information of feature sequence to calculate the posterior probability of phoneme. Then the number of stack recurrent neural network layers in attention model is reduced in order to decrease the calculation of gradient. Experiments in the TIMIT corpus showed that the phoneme error rate is 17.80% in test set, the average training iteration decreased 52%, and the number of training iterations decreased from 139 to 89. The word error rate of WSJ eval92 is 12.9% without any external language model.

The deep learning-based speech enhancement has shown considerable success. However, it still suffers performance degradation under mismatch conditions. In this paper, an adaptation method is proposed to improve the performance under noise mismatch conditions. Firstly, we advise a noise aware training by supplying identity vectors (ivectors) as parallel input features to adapt DNN acoustic models with the target noise. Secondly, given a small amount adaptation data, the noise-dependent DNN is obtained by using Euclidean distance regularization from a noiseindependent DNN, and forcing the estimated masks to be close to the unadapted condition. Finally, experiments were carried out on different noise and SNR conditions, and the proposed method has achieved significantly 29% benefits of STOI at most and provided consistent improvement in PESQ and segSNR against the baseline systems.