Low-frequency word enhancement with similar pairs in speech recognition

In practical automatic speech recognition (ASR) systems, it is difficult to recognize words that are with low-frequency in the language model (LM) training data. Ironically, these words tend to be highly important as they are often domain-specific name entities. In order to meet this challenge, we present a novel approach that enhances the weights of these words by borrowing information from some high-frequency words that are similar to the target words. Experimental results demonstrated that our method can significantly improve ASR performance on low-frequency words and does not impact performance on high-frequency words. Additionally, this method can be easily extended to deal with new words that are absent in the LM training data.