Studies On Speech Recognition Error Detection And Correction Based On Example-Context

Speech recognition is a very important human-computer interaction technology. However, in order to realize the widely practical application, the recognition accuracy rate needs to be further improved. Research on speech recognition error detection and correction by natural language understanding (NLU) method will be an important research direction of improving the performance of speech recognition.In this thesis, an example-context-based approach for speech recognition error detection and correction is proposed, which makes comprehensive use of syntax, semantics, and context information resources, focusing on contribution of the context information, to detect and correct the error of speech recognition text. The main research work and achievements are as follows:1. Research on the representation and calculation of context knowledge. This thesis presents a method of example-based context representation, and computes context knowledge by context correlation rate. On this basis, the author constructed a context knowledge base.2. Research on calculation of sentence correlation. In this thesis, the sentence correlation is a weighted average of word correlation including semantic similarity and context correlation with key word. In addition, words order factor is also taken into account.3. Applying NLU method to post-processing of speech recognition, the thesis designed a speech recognition error detection and correction system based on example-context. The system includes four modules: locating the anchor words, examples extraction, error detection and error correction. The anchor words are located by Phonetic analysis that belongs to syntax analysis, and by key words list that belongs to contextual analysis; examples are extracted by calculating sentence correlation, which belongs to comprehensive analysis of semantics and context; at last, detect and correct the error according to contextual harmony degree in the example-context, and output the optimized result.Through implementation and system test, the feasibility and effectiveness of the method was verified, which improved the precision of speech recognition by 20%.