Linearly Interpolated Hierarchical N-gram Language Models for Speech Recognition Engines

Language modeling is a crucial component in natural language continuous speech recognition, due to the difficulty involved by continuous speech [1], [2]. Language modeling attempts to capture regularities in natural language for the purpose of improving the recognition performance. Many studies have shown that the word error rate of automatic speech recognition (ASR) systems decreases significantly when using statistical language models [3], [4], [5]. The purpose of language models (LMs) is to compute the probability of a sequence of words . The probability can be expressed as: , where is the history or the context of word . The probability becomes difficult to estimate as the number of words in increases. To overcome this problem, we can introduce equivalent classes on the histories in order to reduce their cardinality. The n-gram language models approximate the dependence of each word (regardless of i) to the n — 1 words preceding it: . The probability can then be expressed as: