Research On Tibetan Language Model For Continuous Speech Recognition

The speech recognition technology is to allow the machine through the recognition of the voice signal into a corresponding text. It mainly includes three aspects:feature extraction, pattern matching and model training. Speech recognition technology in present Chinese and English is widely applied language is quite mature, but the application of speech recognition in Tibetan and other minority languages are few. The purpose of this paper is to study the application of language model in Tibetan continuous speech recognition system. Speech recognition includes acoustic and language model, relative to the acoustic model to study the mature, language model has a more broad prospects for development. Experiments show that, the study of language model to improve the speech recognition rate has greatly improved.The Tibetan words of Lhasa, put forward the new text expected screening programs. The use of HTK tools for processing of data, generated dictionary, generative language model. Due to the sparse data problem, statistical language model smoothing algorithm directly affects the performance of speech recognition systems. This paper detailed analysis and comparison of the additive smoothing algorithm, back off smoothing algorithm, linear interpolation smoothing algorithm and nonlinear discount smoothing algorithm. Through the identification and comparison of various algorithms confusion degree and rate of speech recognition system, and ultimately selected Kneser-Ney smoothing algorithm is revised to HTK language based on continuous speech recognition system.Corpus used in this article is China Minorities Information Technology Institute at Northwestern University recorded300,000ethnic Tibetan Lhasa dialect news corpus, established in accordance with one of the most commonly used dictionary5700syllables produce ternary statistical language model through training. Experimental results show that the various smoothing algorithm, the modified version of Kneser-Ney smoothing algorithm that confusion is minimum, the application of this sentence recognition rate smoothing algorithm Tibetan continuous speech recognition system from the original67.83%to78.98%, word recognition rate increased from91.34%up to94.78%.