ASR Hypothesis Reranking Using Prior-Informed Restricted Boltzmann Machine

Abstract

Discriminative language models (DLMs) have been widely used for reranking competing hypotheses produced by an Automatic Speech Recognition (ASR) system. While existing DLMs suffer from limited generalization power, we propose a novel DLM based on a discriminatively trained Restricted Boltzmann Machine (RBM). The hidden layer of the RBM improves generalization and allows for employing additional prior knowledge, including pre-trained parameters and entity-related prior. Our approach outperforms the single-layer-perceptron (SLP) reranking model, and fusing our approach with SLP achieves up to 1.3% absolute Word Error Rate (WER) reduction and a relative 180% improvement in terms of WER reduction over the SLP reranker. In particular, it shows that proposed prior informed RBM reranker achieves largest ASR error reduction (3.1% absolute WER) on content words.