Large-Context Automatic Speech Recognition Based on RNN Transducer

We propose large-context end-to-end automatic speech recognition (ASR) based on a recurrent neural network transducer (RNN-T) for conversational ASR. Typical end-to-end ASR models recognize independent utterances. Unlike the typical end-to-end ASR models, attention-based encoder-decoder models, which utilize a large context (i.e., a long-range context beyond the boundary of an utterance) have been recently proposed for conversational ASR. In this work, we introduce a large-context encoder for RNN-T to utilize hypotheses generated in previous utterances as a large context. The large-context encoder obtains concatenated hypotheses and converts them to hidden vectors. The hidden vectors are summarized as the available context based on an attention mechanism. During training, the large-context encoder and RNN-T components are jointly optimized using RNN-T loss. In our experiments, we evaluate the proposed method using medical conversations as a dataset. As a result of the experiment, we obtained 17.2 % and 6.0 % relative reductions in the character error rate compared with those of the original RNN-T on utterances spoken by the doctor and patient, respectively.