Automatic Construction of FSA Language Model for Speech Recognition by FSA DP-Matching

For accurate speech recognition, a well-defined language model is necessary. When a large amount of learning texts (corpus) is available, a statistical language model such as bi-gram or tri-gram generated from a corpus is quite powerful. For example, most of current dictation systems employ bi-gram or tri-gram language models. However, if the size of the corpus is not sufficiently large, reliability of statistical information calculated from the corpus decreases and then so does the effectiveness of the generated statistical language model (sparseness problem). Furthermore, preparing a sufficient amount of texts for spoken language is generally very expensive. Therefore, a finite state automaton (FSA) language model is generally used for small- or middle-size (around 1000 words) vocabulary speech recognition. However, defining a FSA model by hand requires much human effort. In some systems, a FSA model is automatically generated by converting from regular-grammar type grammar rules (e.g., see Hparse [1]), but it is still a very time- and effort-consuming task to prepare a grammar with sufficient coverage and consistency.

In this chapter, we propose a new method to construct a FSA language model, by using a FSA DP (dynamic programming) matching method. We also report some experimental results; the method was applied to a travel conversation corpus (with about a thousand sentences) to generate a FSA model, and then, speech recognition experiments using the language model were conducted. The result shows that the recognition correct rate is high enough for closed data, but is not so satisfactory for open data. To cope with this problem, we also propose an additional mechanism that decides whether a recognized result can be accepted or should be rejected by evaluating a distance of the result from the learning corpus. After the mechanism is applied, the recognition correct rate for accepted results is considerably improved.