A Speech Recognition Acoustic Model Based on LSTM -CTC

Acoustic model plays a very important role in the voice recognition systems. Compared with most of the previous systems which using discriminant models combined with HMM hybrid model for acoustic model training, this paper proposes a LSTM-CTC model by combining CTC training with LSTM model based on the principle of Connectionist Temporal Classification (CTC). By inserting the Softmax vector output from the top of LSTM into the CTC model and using the CTC decoding method, this model reduces the loss of the entire sequence and predicts the sequence label correctly in the prediction probability of the LSTM output. Experiments show that the new model that compared with DNN-HMM model not only reduces the word error rate by 1.4 percentage points, but also shortens the model training time by nearly three-quarters.