CNN-Based Audio Front End Processing on Speech Recognition

Recently, Convolutional Neural Networks (CNNs) have become widely used in the field of speech recognition. Their role is to mitigate the spectral shifts caused by differences between the speakers and the environment. However, how to better use CNN for speech recognition has always been a problem worth exploring. This paper study the usage of CNN as the audio front end processing network based on the acoustic model architecture of Long Short-Term Memory and Connectionist Temporal Classification (LSTM-CTC). Firstly, by comparing the different inputs of the model, a Full-Mel Spectrogram feature that accords with the auditory characteristics is proposed. After determining the input features, this paper explores the principles of CNN parameters design and its role in speech recognition through the comparative experiments of different parameters. Finally, when decoding with phoneme-level language model, we achieve a test set error of 16.9% on the TIMIT phoneme recognition benchmark, which is the best known result of using the CTC as loss function.