A Novel Spec-CNN-CTC Model for End-to-End Speech Recognition

This paper discusses the application of a special data augmentation approach for end-to-end phone recognition system on the Deep Neural Networks. The system improves the performance of phone recognition and alleviates overfitting during training. Also, it offers a solution to the problem of few public datasets annotated at the phone level. And we propose the CNN-CTC structure as a baseline model. The model is based on Convolutional Neural Networks (CNNs) and Connectionist Temporal Classification (CTC) objective function. Which is an end-to-end structure, and there is no need to force alignment each frame of audio. The SpecAugment approach directly processes the feature of audio, such as the log Mel-spectrogram. In our experiment, the Spec-CNN-CTC system achieves a phone error rate of 16.11% on TIMIT corpus with no prior linguistic information. Which is outperforming the previous work Acoustic-State-Transition Model (ASTM) by 27.63%, the DNN-HMM with MFCC + IFCC features by 16.8%, the RNN-CRF model by 17.3% and the DBM-DNN model by 22.62%.