Hybrid architecture CNN-BLSTM for automatic speech recognition

With several deep learning approaches, the domain of automatic speech recognition (ASR) has seen notable advancements in recent times. The domains of intelligent human- computer interaction and machine translation greatly benefited from accurate speech recognition. The present study introduces a hybrid architecture amalgamating a convolutional neural network (CNN) and bidirectional long short-term memory (BLSTM) for speech recognition. For aligning speech input sequences with corresponding text output sequences, it uses the connectionist temporal classification (CTC) technique. The experiments were done on the LJ speech dataset, and the results showed with an increased number of training samples, the performance of the speech recognition algorithm tended to be increased but the time taken for training gradually increased over time. Moreover, a trained speech recognition algorithm exhibits a longer training time when the recognition accuracy is lower. In this study, we implemented a hybrid deep learning model CNN-BLSTM, in conjunction with the CTC loss function, which attains a word error rate (WER) of 36.97% on the testing dataset.