PAT-tree-based Language Modeling with Initial Application of Chinese Speech Recognition Output Verification

In spontaneous speech recognition, there are always inevitable errors in the output due to the difficulties of acoustic recognition or linguistic decoding. In this paper, we present an output verification approach to detect and correct the errors automatically using the abundant Internet resources. The Syllable PAT tree (SPAT tree), a metamorphic data structure derived from the PAT tree concept, is a real N-gram language model and is first used as a verifier for speech recognition output in order to improve the accuracy of speech recognition. The verification approaches proposed here not only reduce the character error rate by 12.66% in preliminary experiments, but can make the recognition results more reliable for the following-up processing, such as semantic analysis in dialog control or speech understanding.