Research On Error Handling Of The Speech Recogition Post-Processing

After the speech recognition processing error detection and error correction is one of the indispensable step in the process of speech understanding.Error detection and error correction are usually based on the decoder output information and high-level language semantics information that is separated with the speech recognition system. The confusion matrix is also commonly used in the error detection and error correction being.This article is based on the simple confusion matrix, make the rich statistical information to guide the error detection and error correction processes.The main work and innovation of this paper is as follows:Firstly, the paper do some research on the confusion network based weighted confusion matrix algorithms.This paper relies on the use of the confusion matrix, and so accurate and detailed confusion matrix plays a crucial role.Based on the proposed confusion matrix generation algorithm, work is done on the prospects that different acoustic units of the same confusion set makes a distinction between statistical weighting.Experiments showed that this improved method does well in the subsequent error correction and error detection processes.Secondly, the paper studies the algorithm of speech recognition errors correction based on confusion matrix.Based on confusion matrix, this paper proposes the idea of calculating the center of the candidates space to deal with the error correction problems.The different candidates in the same set have close relationship, they originate the same acoustic unit.In order to quantify the relationship, confusion vector is creatively used in the representation of the candidates, and the problem of how to calculate the center of the candidates space is introduced.Thus the error correction problem gives a measure of the high-dimensional space, the experiments proves this using statistical information for error correction method really do some value.Thirdly, confusion-term-fixalgorithm based on similarity graph is proposed to deal with the error detection and correction problems.Further research is done based on the algorithm proposed above, and confusion-term-fix algorithm is proposed. Confusion matrix has rich experience and statistical information, which is clearly beneficial to guide the error correction process.But the real object of error correction is a result of recognition of specific events, which has a certain distance to the statistical experiences.The graph algorithm is creatively used to do some balance, which plays a significant role. The experiments proves this algorithm really do some value.