Abstract: With the development of big data, speech recognition technology based on deep learning has been quite mature, but under small sample resources, due to the limited relevance of feature information, the modeling ability of contextual information of the model is insufficient, which leads to low recognition rate. To solve this problem, a timing prediction acoustic model (named TLSTM-Attention), which consists of a time delay neural network (TDNN) embedded by attention mechanism layer (Attention) and a long and short time memory (LSTM) recurrent neural network, is proposed in this paper. This model can effectively fuse the coarse and fine particle features with important information to improve the modeling ability of context information. By using the velocity perturbation technique to amplify the data and combining the speaker's channel information features and the lattice-free maximum mutual information training criteria, and by selecting different input features, model structures and numbers of nodes, a series of comparative experiments are conducted. The experimental results show that compared with the baseline model, the word error rate of the model is reduced by 3.77 percentage points.