LSTM-Based Language Models for Very Large Vocabulary Continuous Russian Speech Recognition System

摘要

This paper presents language models based on Long Short-Term Memory (LSTM) neural networks for very large vocabulary continuous Russian speech recognition. We created neural networks with various numbers of units in hidden and projection layers using different optimization methods. Obtained LSTM-based language models were used for N-best list rescoring. As well we tested a linear interpolation of LSTM language model with the baseline 3-gram language model and achieved 22% relative reduction of the word error rate with respect to the baseline 3-gram model.