Language Model Integration Based on Memory Control for Sequence to Sequence Speech Recognition

In this paper, we explore several new schemes to train a seq2seqmodel to integrate a pre-trained language model (LM). Our proposedfusion methods focus on the memory cell state and the hidden statein the seq2seq decoder long short-term memory (LSTM), and thememory cell state is updated by the LM unlike the prior studies.This means the memory retained by the main seq2seq would be adjustedby the external LM. These fusion methods have several variantsdepending on the architecture of this memory cell update andthe use of memory cell and hidden states which directly affects thefinal label inference. We performed the experiments to show the effectivenessof the proposed methods in a mono-lingual ASR setup onthe Librispeech corpus and in a transfer learning setup from a multilingualASR (MLASR) base model to a low-resourced language. InLibrispeech, our best model improved WER by 3.7%, 2.4% for testclean, test other relatively to the shallow fusion baseline, with multileveldecoding. In transfer learning from an MLASR base modelto the IARPA Babel Swahili model, the best scheme improved thetransferred model on eval set by 9.9%, 9.8% in CER, WER relativelyto the 2-stage transfer baseline.