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

Abstract

In this paper, we explore several new schemes to train a seq2seq model to integrate a pre-trained LM. Our proposed fusion methods focus on the memory cell state and the hidden state in the seq2seq decoder long short-term memory (LSTM), and the memory cell state is updated by the LM unlike the prior studies. This means the memory retained by the main seq2seq would be adjusted by the external LM. These fusion methods have several variants depending on the architecture of this memory cell update and the use of memory cell and hidden states which directly affects the final label inference. We performed the experiments to show the effectiveness of the proposed methods in a mono-lingual ASR setup on the Librispeech corpus and in a transfer learning setup from a multilingual ASR (MLASR) base model to a low-resourced language. In Librispeech, our best model improved WER by 3.7%, 2.4% for test clean, test other relatively to the shallow fusion baseline, with multi-level decoding. In transfer learning from an MLASR base model to the IARPA Babel Swahili model, the best scheme improved the transferred model on eval set by 9.9%, 9.8% in CER, WER relatively to the 2-stage transfer baseline.

Text Augmentation for Language Models in High Error Recognition Scenario

“…For all augmentation schemes, we sweep across the rate of dropout in range [0.0, 0.7] to find the optimal level of total regularization. For the baseline, the best result comes from setting it to 0.7 3 , those trained with data augmentation were fairly robust to the dropout rate and achieved their best performance in range of 0.3 -0.6.…”

“…The traditional reason language models (LMs) appear in ASR systems is that they directly represent the prior term P (S) in the Bayes factorization of the posterior probability P (S|A) of a sentence S given the audio A. However in practice, LMs trained on excessive amounts of data are combined with hybrid and end-to-end systems alike [1,2,3] at authors liberty. Overall, LMs can be seen as a refinement tool to apply on a preliminary result of recognition.…”

A Survey on Knowledge integration techniques with Artificial Neural Networks for seq-2-seq/time series models

“…Cho et al [9] presents a technique (Cell Control Fusion) that is similar to [8], but differs in the aspect of not just fusing the gated outputs of hidden states of the external language model (RNN) but also for the cell states. Hence, a LSTM flavor of the RNN is used as a sequence-to-sequence model in this technique.…”

Fusion Models for Improved Visual Captioning

“…Unimodal and multimodal model fusion has been explored extensively in the context of ASR [29,7], Neural Machine Translation (NMT) [12], and hierarchical story generation [11]. However, to the best of our knowledge, there has been no similar works for visual captioning.…”